fs/jfs/jfs_logmgr.c

   1 /*
   2  *   Copyright (C) International Business Machines Corp., 2000-2004
   3  *   Portions Copyright (C) Christoph Hellwig, 2001-2002
   4  *
   5  *   This program is free software;  you can redistribute it and/or modify
   6  *   it under the terms of the GNU General Public License as published by
   7  *   the Free Software Foundation; either version 2 of the License, or
   8  *   (at your option) any later version.
   9  *
  10  *   This program is distributed in the hope that it will be useful,
  11  *   but WITHOUT ANY WARRANTY;  without even the implied warranty of
  12  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See
  13  *   the GNU General Public License for more details.
  14  *
  15  *   You should have received a copy of the GNU General Public License
  16  *   along with this program;  if not, write to the Free Software
  17  *   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  18  */
  19
  20 /*
  21  *      jfs_logmgr.c: log manager
  22  *
  23  * for related information, see transaction manager (jfs_txnmgr.c), and
  24  * recovery manager (jfs_logredo.c).
  25  *
  26  * note: for detail, RTFS.
  27  *
  28  *      log buffer manager:
  29  * special purpose buffer manager supporting log i/o requirements.
  30  * per log serial pageout of logpage
  31  * queuing i/o requests and redrive i/o at iodone
  32  * maintain current logpage buffer
  33  * no caching since append only
  34  * appropriate jfs buffer cache buffers as needed
  35  *
  36  *      group commit:
  37  * transactions which wrote COMMIT records in the same in-memory
  38  * log page during the pageout of previous/current log page(s) are
  39  * committed together by the pageout of the page.
  40  *
  41  *      TBD lazy commit:
  42  * transactions are committed asynchronously when the log page
  43  * containing it COMMIT is paged out when it becomes full;
  44  *
  45  *      serialization:
  46  * . a per log lock serialize log write.
  47  * . a per log lock serialize group commit.
  48  * . a per log lock serialize log open/close;
  49  *
  50  *      TBD log integrity:
  51  * careful-write (ping-pong) of last logpage to recover from crash
  52  * in overwrite.
  53  * detection of split (out-of-order) write of physical sectors
  54  * of last logpage via timestamp at end of each sector
  55  * with its mirror data array at trailer).
  56  *
  57  *      alternatives:
  58  * lsn - 64-bit monotonically increasing integer vs
  59  * 32-bit lspn and page eor.
  60  */
  61
  62 #include <linux/fs.h>
  63 #include <linux/blkdev.h>
  64 #include <linux/interrupt.h>
  65 #include <linux/smp_lock.h>
  66 #include <linux/completion.h>
  67 #include <linux/buffer_head.h>          /* for sync_blockdev() */
  68 #include <linux/bio.h>
  69 #include <linux/suspend.h>
  70 #include "jfs_incore.h"
  71 #include "jfs_filsys.h"
  72 #include "jfs_metapage.h"
  73 #include "jfs_txnmgr.h"
  74 #include "jfs_debug.h"
  75
  76
  77 /*
  78  * lbuf's ready to be redriven.  Protected by log_redrive_lock (jfsIO thread)
  79  */
  80 static struct lbuf *log_redrive_list;
  81 static spinlock_t log_redrive_lock = SPIN_LOCK_UNLOCKED;
  82 DECLARE_WAIT_QUEUE_HEAD(jfs_IO_thread_wait);
  83
  84
  85 /*
  86  *      log read/write serialization (per log)
  87  */
  88 #define LOG_LOCK_INIT(log)      init_MUTEX(&(log)->loglock)
  89 #define LOG_LOCK(log)           down(&((log)->loglock))
  90 #define LOG_UNLOCK(log)         up(&((log)->loglock))
  91
  92
  93 /*
  94  *      log group commit serialization (per log)
  95  */
  96
  97 #define LOGGC_LOCK_INIT(log)    spin_lock_init(&(log)->gclock)
  98 #define LOGGC_LOCK(log)         spin_lock_irq(&(log)->gclock)
  99 #define LOGGC_UNLOCK(log)       spin_unlock_irq(&(log)->gclock)
 100 #define LOGGC_WAKEUP(tblk)      wake_up_all(&(tblk)->gcwait)
 101
 102 /*
 103  *      log sync serialization (per log)
 104  */
 105 #define LOGSYNC_DELTA(logsize)          min((logsize)/8, 128*LOGPSIZE)
 106 #define LOGSYNC_BARRIER(logsize)        ((logsize)/4)
 107 /*
 108 #define LOGSYNC_DELTA(logsize)          min((logsize)/4, 256*LOGPSIZE)
 109 #define LOGSYNC_BARRIER(logsize)        ((logsize)/2)
 110 */
 111
 112
 113 /*
 114  *      log buffer cache synchronization
 115  */
 116 static spinlock_t jfsLCacheLock = SPIN_LOCK_UNLOCKED;
 117
 118 #define LCACHE_LOCK(flags)      spin_lock_irqsave(&jfsLCacheLock, flags)
 119 #define LCACHE_UNLOCK(flags)    spin_unlock_irqrestore(&jfsLCacheLock, flags)
 120
 121 /*
 122  * See __SLEEP_COND in jfs_locks.h
 123  */
 124 #define LCACHE_SLEEP_COND(wq, cond, flags)      \
 125 do {                                            \
 126         if (cond)                               \
 127                 break;                          \
 128         __SLEEP_COND(wq, cond, LCACHE_LOCK(flags), LCACHE_UNLOCK(flags)); \
 129 } while (0)
 130
 131 #define LCACHE_WAKEUP(event)    wake_up(event)
 132
 133
 134 /*
 135  *      lbuf buffer cache (lCache) control
 136  */
 137 /* log buffer manager pageout control (cumulative, inclusive) */
 138 #define lbmREAD         0x0001
 139 #define lbmWRITE        0x0002  /* enqueue at tail of write queue;
 140                                  * init pageout if at head of queue;
 141                                  */
 142 #define lbmRELEASE      0x0004  /* remove from write queue
 143                                  * at completion of pageout;
 144                                  * do not free/recycle it yet:
 145                                  * caller will free it;
 146                                  */
 147 #define lbmSYNC         0x0008  /* do not return to freelist
 148                                  * when removed from write queue;
 149                                  */
 150 #define lbmFREE         0x0010  /* return to freelist
 151                                  * at completion of pageout;
 152                                  * the buffer may be recycled;
 153                                  */
 154 #define lbmDONE         0x0020
 155 #define lbmERROR        0x0040
 156 #define lbmGC           0x0080  /* lbmIODone to perform post-GC processing
 157                                  * of log page
 158                                  */
 159 #define lbmDIRECT       0x0100
 160
 161 /*
 162  * Global list of active external journals
 163  */
 164 static LIST_HEAD(jfs_external_logs);
 165 static struct jfs_log *dummy_log = NULL;
 166 static DECLARE_MUTEX(jfs_log_sem);
 167
 168 /*
 169  * external references
 170  */
 171 extern void txLazyUnlock(struct tblock * tblk);
 172 extern int jfs_stop_threads;
 173 extern struct completion jfsIOwait;
 174 extern int jfs_tlocks_low;
 175
 176 /*
 177  * forward references
 178  */
 179 static int lmWriteRecord(struct jfs_log * log, struct tblock * tblk,
 180                          struct lrd * lrd, struct tlock * tlck);
 181
 182 static int lmNextPage(struct jfs_log * log);
 183 static int lmLogFileSystem(struct jfs_log * log, struct jfs_sb_info *sbi,
 184                            int activate);
 185
 186 static int open_inline_log(struct super_block *sb);
 187 static int open_dummy_log(struct super_block *sb);
 188 static int lbmLogInit(struct jfs_log * log);
 189 static void lbmLogShutdown(struct jfs_log * log);
 190 static struct lbuf *lbmAllocate(struct jfs_log * log, int);
 191 static void lbmFree(struct lbuf * bp);
 192 static void lbmfree(struct lbuf * bp);
 193 static int lbmRead(struct jfs_log * log, int pn, struct lbuf ** bpp);
 194 static void lbmWrite(struct jfs_log * log, struct lbuf * bp, int flag, int cant_block);
 195 static void lbmDirectWrite(struct jfs_log * log, struct lbuf * bp, int flag);
 196 static int lbmIOWait(struct lbuf * bp, int flag);
 197 static bio_end_io_t lbmIODone;
 198 static void lbmStartIO(struct lbuf * bp);
 199 static void lmGCwrite(struct jfs_log * log, int cant_block);
 200 static int lmLogSync(struct jfs_log * log, int nosyncwait);
 201
 202
 203
 204 /*
 205  *      statistics
 206  */
 207 #ifdef CONFIG_JFS_STATISTICS
 208 static struct lmStat {
 209         uint commit;            /* # of commit */
 210         uint pagedone;          /* # of page written */
 211         uint submitted;         /* # of pages submitted */
 212         uint full_page;         /* # of full pages submitted */
 213         uint partial_page;      /* # of partial pages submitted */
 214 } lmStat;
 215 #endif
 216
 217
 218 /*
 219  * NAME:        lmLog()
 220  *
 221  * FUNCTION:    write a log record;
 222  *
 223  * PARAMETER:
 224  *
 225  * RETURN:      lsn - offset to the next log record to write (end-of-log);
 226  *              -1  - error;
 227  *
 228  * note: todo: log error handler
 229  */
 230 int lmLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
 231           struct tlock * tlck)
 232 {
 233         int lsn;
 234         int diffp, difft;
 235         struct metapage *mp = NULL;
 236
 237         jfs_info("lmLog: log:0x%p tblk:0x%p, lrd:0x%p tlck:0x%p",
 238                  log, tblk, lrd, tlck);
 239
 240         LOG_LOCK(log);
 241
 242         /* log by (out-of-transaction) JFS ? */
 243         if (tblk == NULL)
 244                 goto writeRecord;
 245
 246         /* log from page ? */
 247         if (tlck == NULL ||
 248             tlck->type & tlckBTROOT || (mp = tlck->mp) == NULL)
 249                 goto writeRecord;
 250
 251         /*
 252          *      initialize/update page/transaction recovery lsn
 253          */
 254         lsn = log->lsn;
 255
 256         LOGSYNC_LOCK(log);
 257
 258         /*
 259          * initialize page lsn if first log write of the page
 260          */
 261         if (mp->lsn == 0) {
 262                 mp->log = log;
 263                 mp->lsn = lsn;
 264                 log->count++;
 265
 266                 /* insert page at tail of logsynclist */
 267                 list_add_tail(&mp->synclist, &log->synclist);
 268         }
 269
 270         /*
 271          *      initialize/update lsn of tblock of the page
 272          *
 273          * transaction inherits oldest lsn of pages associated
 274          * with allocation/deallocation of resources (their
 275          * log records are used to reconstruct allocation map
 276          * at recovery time: inode for inode allocation map,
 277          * B+-tree index of extent descriptors for block
 278          * allocation map);
 279          * allocation map pages inherit transaction lsn at
 280          * commit time to allow forwarding log syncpt past log
 281          * records associated with allocation/deallocation of
 282          * resources only after persistent map of these map pages
 283          * have been updated and propagated to home.
 284          */
 285         /*
 286          * initialize transaction lsn:
 287          */
 288         if (tblk->lsn == 0) {
 289                 /* inherit lsn of its first page logged */
 290                 tblk->lsn = mp->lsn;
 291                 log->count++;
 292
 293                 /* insert tblock after the page on logsynclist */
 294                 list_add(&tblk->synclist, &mp->synclist);
 295         }
 296         /*
 297          * update transaction lsn:
 298          */
 299         else {
 300                 /* inherit oldest/smallest lsn of page */
 301                 logdiff(diffp, mp->lsn, log);
 302                 logdiff(difft, tblk->lsn, log);
 303                 if (diffp < difft) {
 304                         /* update tblock lsn with page lsn */
 305                         tblk->lsn = mp->lsn;
 306
 307                         /* move tblock after page on logsynclist */
 308                         list_move(&tblk->synclist, &mp->synclist);
 309                 }
 310         }
 311
 312         LOGSYNC_UNLOCK(log);
 313
 314         /*
 315          *      write the log record
 316          */
 317       writeRecord:
 318         lsn = lmWriteRecord(log, tblk, lrd, tlck);
 319
 320         /*
 321          * forward log syncpt if log reached next syncpt trigger
 322          */
 323         logdiff(diffp, lsn, log);
 324         if (diffp >= log->nextsync)
 325                 lsn = lmLogSync(log, 0);
 326
 327         /* update end-of-log lsn */
 328         log->lsn = lsn;
 329
 330         LOG_UNLOCK(log);
 331
 332         /* return end-of-log address */
 333         return lsn;
 334 }
 335
 336
 337 /*
 338  * NAME:        lmWriteRecord()
 339  *
 340  * FUNCTION:    move the log record to current log page
 341  *
 342  * PARAMETER:   cd      - commit descriptor
 343  *
 344  * RETURN:      end-of-log address
 345  *
 346  * serialization: LOG_LOCK() held on entry/exit
 347  */
 348 static int
 349 lmWriteRecord(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
 350               struct tlock * tlck)
 351 {
 352         int lsn = 0;            /* end-of-log address */
 353         struct lbuf *bp;        /* dst log page buffer */
 354         struct logpage *lp;     /* dst log page */
 355         caddr_t dst;            /* destination address in log page */
 356         int dstoffset;          /* end-of-log offset in log page */
 357         int freespace;          /* free space in log page */
 358         caddr_t p;              /* src meta-data page */
 359         caddr_t src;
 360         int srclen;
 361         int nbytes;             /* number of bytes to move */
 362         int i;
 363         int len;
 364         struct linelock *linelock;
 365         struct lv *lv;
 366         struct lvd *lvd;
 367         int l2linesize;
 368
 369         len = 0;
 370
 371         /* retrieve destination log page to write */
 372         bp = (struct lbuf *) log->bp;
 373         lp = (struct logpage *) bp->l_ldata;
 374         dstoffset = log->eor;
 375
 376         /* any log data to write ? */
 377         if (tlck == NULL)
 378                 goto moveLrd;
 379
 380         /*
 381          *      move log record data
 382          */
 383         /* retrieve source meta-data page to log */
 384         if (tlck->flag & tlckPAGELOCK) {
 385                 p = (caddr_t) (tlck->mp->data);
 386                 linelock = (struct linelock *) & tlck->lock;
 387         }
 388         /* retrieve source in-memory inode to log */
 389         else if (tlck->flag & tlckINODELOCK) {
 390                 if (tlck->type & tlckDTREE)
 391                         p = (caddr_t) &JFS_IP(tlck->ip)->i_dtroot;
 392                 else
 393                         p = (caddr_t) &JFS_IP(tlck->ip)->i_xtroot;
 394                 linelock = (struct linelock *) & tlck->lock;
 395         }
 396 #ifdef  _JFS_WIP
 397         else if (tlck->flag & tlckINLINELOCK) {
 398
 399                 inlinelock = (struct inlinelock *) & tlck;
 400                 p = (caddr_t) & inlinelock->pxd;
 401                 linelock = (struct linelock *) & tlck;
 402         }
 403 #endif                          /* _JFS_WIP */
 404         else {
 405                 jfs_err("lmWriteRecord: UFO tlck:0x%p", tlck);
 406                 return 0;       /* Probably should trap */
 407         }
 408         l2linesize = linelock->l2linesize;
 409
 410       moveData:
 411         ASSERT(linelock->index <= linelock->maxcnt);
 412
 413         lv = linelock->lv;
 414         for (i = 0; i < linelock->index; i++, lv++) {
 415                 if (lv->length == 0)
 416                         continue;
 417
 418                 /* is page full ? */
 419                 if (dstoffset >= LOGPSIZE - LOGPTLRSIZE) {
 420                         /* page become full: move on to next page */
 421                         lmNextPage(log);
 422
 423                         bp = log->bp;
 424                         lp = (struct logpage *) bp->l_ldata;
 425                         dstoffset = LOGPHDRSIZE;
 426                 }
 427
 428                 /*
 429                  * move log vector data
 430                  */
 431                 src = (u8 *) p + (lv->offset << l2linesize);
 432                 srclen = lv->length << l2linesize;
 433                 len += srclen;
 434                 while (srclen > 0) {
 435                         freespace = (LOGPSIZE - LOGPTLRSIZE) - dstoffset;
 436                         nbytes = min(freespace, srclen);
 437                         dst = (caddr_t) lp + dstoffset;
 438                         memcpy(dst, src, nbytes);
 439                         dstoffset += nbytes;
 440
 441                         /* is page not full ? */
 442                         if (dstoffset < LOGPSIZE - LOGPTLRSIZE)
 443                                 break;
 444
 445                         /* page become full: move on to next page */
 446                         lmNextPage(log);
 447
 448                         bp = (struct lbuf *) log->bp;
 449                         lp = (struct logpage *) bp->l_ldata;
 450                         dstoffset = LOGPHDRSIZE;
 451
 452                         srclen -= nbytes;
 453                         src += nbytes;
 454                 }
 455
 456                 /*
 457                  * move log vector descriptor
 458                  */
 459                 len += 4;
 460                 lvd = (struct lvd *) ((caddr_t) lp + dstoffset);
 461                 lvd->offset = cpu_to_le16(lv->offset);
 462                 lvd->length = cpu_to_le16(lv->length);
 463                 dstoffset += 4;
 464                 jfs_info("lmWriteRecord: lv offset:%d length:%d",
 465                          lv->offset, lv->length);
 466         }
 467
 468         if ((i = linelock->next)) {
 469                 linelock = (struct linelock *) lid_to_tlock(i);
 470                 goto moveData;
 471         }
 472
 473         /*
 474          *      move log record descriptor
 475          */
 476       moveLrd:
 477         lrd->length = cpu_to_le16(len);
 478
 479         src = (caddr_t) lrd;
 480         srclen = LOGRDSIZE;
 481
 482         while (srclen > 0) {
 483                 freespace = (LOGPSIZE - LOGPTLRSIZE) - dstoffset;
 484                 nbytes = min(freespace, srclen);
 485                 dst = (caddr_t) lp + dstoffset;
 486                 memcpy(dst, src, nbytes);
 487
 488                 dstoffset += nbytes;
 489                 srclen -= nbytes;
 490
 491                 /* are there more to move than freespace of page ? */
 492                 if (srclen)
 493                         goto pageFull;
 494
 495                 /*
 496                  * end of log record descriptor
 497                  */
 498
 499                 /* update last log record eor */
 500                 log->eor = dstoffset;
 501                 bp->l_eor = dstoffset;
 502                 lsn = (log->page << L2LOGPSIZE) + dstoffset;
 503
 504                 if (lrd->type & cpu_to_le16(LOG_COMMIT)) {
 505                         tblk->clsn = lsn;
 506                         jfs_info("wr: tclsn:0x%x, beor:0x%x", tblk->clsn,
 507                                  bp->l_eor);
 508
 509                         INCREMENT(lmStat.commit);       /* # of commit */
 510
 511                         /*
 512                          * enqueue tblock for group commit:
 513                          *
 514                          * enqueue tblock of non-trivial/synchronous COMMIT
 515                          * at tail of group commit queue
 516                          * (trivial/asynchronous COMMITs are ignored by
 517                          * group commit.)
 518                          */
 519                         LOGGC_LOCK(log);
 520
 521                         /* init tblock gc state */
 522                         tblk->flag = tblkGC_QUEUE;
 523                         tblk->bp = log->bp;
 524                         tblk->pn = log->page;
 525                         tblk->eor = log->eor;
 526
 527                         /* enqueue transaction to commit queue */
 528                         list_add_tail(&tblk->cqueue, &log->cqueue);
 529
 530                         LOGGC_UNLOCK(log);
 531                 }
 532
 533                 jfs_info("lmWriteRecord: lrd:0x%04x bp:0x%p pn:%d eor:0x%x",
 534                         le16_to_cpu(lrd->type), log->bp, log->page, dstoffset);
 535
 536                 /* page not full ? */
 537                 if (dstoffset < LOGPSIZE - LOGPTLRSIZE)
 538                         return lsn;
 539
 540               pageFull:
 541                 /* page become full: move on to next page */
 542                 lmNextPage(log);
 543
 544                 bp = (struct lbuf *) log->bp;
 545                 lp = (struct logpage *) bp->l_ldata;
 546                 dstoffset = LOGPHDRSIZE;
 547                 src += nbytes;
 548         }
 549
 550         return lsn;
 551 }
 552
 553
 554 /*
 555  * NAME:        lmNextPage()
 556  *
 557  * FUNCTION:    write current page and allocate next page.
 558  *
 559  * PARAMETER:   log
 560  *
 561  * RETURN:      0
 562  *
 563  * serialization: LOG_LOCK() held on entry/exit
 564  */
 565 static int lmNextPage(struct jfs_log * log)
 566 {
 567         struct logpage *lp;
 568         int lspn;               /* log sequence page number */
 569         int pn;                 /* current page number */
 570         struct lbuf *bp;
 571         struct lbuf *nextbp;
 572         struct tblock *tblk;
 573
 574         /* get current log page number and log sequence page number */
 575         pn = log->page;
 576         bp = log->bp;
 577         lp = (struct logpage *) bp->l_ldata;
 578         lspn = le32_to_cpu(lp->h.page);
 579
 580         LOGGC_LOCK(log);
 581
 582         /*
 583          *      write or queue the full page at the tail of write queue
 584          */
 585         /* get the tail tblk on commit queue */
 586         if (list_empty(&log->cqueue))
 587                 tblk = NULL;
 588         else
 589                 tblk = list_entry(log->cqueue.prev, struct tblock, cqueue);
 590
 591         /* every tblk who has COMMIT record on the current page,
 592          * and has not been committed, must be on commit queue
 593          * since tblk is queued at commit queueu at the time
 594          * of writing its COMMIT record on the page before
 595          * page becomes full (even though the tblk thread
 596          * who wrote COMMIT record may have been suspended
 597          * currently);
 598          */
 599
 600         /* is page bound with outstanding tail tblk ? */
 601         if (tblk && tblk->pn == pn) {
 602                 /* mark tblk for end-of-page */
 603                 tblk->flag |= tblkGC_EOP;
 604
 605                 if (log->cflag & logGC_PAGEOUT) {
 606                         /* if page is not already on write queue,
 607                          * just enqueue (no lbmWRITE to prevent redrive)
 608                          * buffer to wqueue to ensure correct serial order
 609                          * of the pages since log pages will be added
 610                          * continuously
 611                          */
 612                         if (bp->l_wqnext == NULL)
 613                                 lbmWrite(log, bp, 0, 0);
 614                 } else {
 615                         /*
 616                          * No current GC leader, initiate group commit
 617                          */
 618                         log->cflag |= logGC_PAGEOUT;
 619                         lmGCwrite(log, 0);
 620                 }
 621         }
 622         /* page is not bound with outstanding tblk:
 623          * init write or mark it to be redriven (lbmWRITE)
 624          */
 625         else {
 626                 /* finalize the page */
 627                 bp->l_ceor = bp->l_eor;
 628                 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
 629                 lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmFREE, 0);
 630         }
 631         LOGGC_UNLOCK(log);
 632
 633         /*
 634          *      allocate/initialize next page
 635          */
 636         /* if log wraps, the first data page of log is 2
 637          * (0 never used, 1 is superblock).
 638          */
 639         log->page = (pn == log->size - 1) ? 2 : pn + 1;
 640         log->eor = LOGPHDRSIZE; /* ? valid page empty/full at logRedo() */
 641
 642         /* allocate/initialize next log page buffer */
 643         nextbp = lbmAllocate(log, log->page);
 644         nextbp->l_eor = log->eor;
 645         log->bp = nextbp;
 646
 647         /* initialize next log page */
 648         lp = (struct logpage *) nextbp->l_ldata;
 649         lp->h.page = lp->t.page = cpu_to_le32(lspn + 1);
 650         lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE);
 651
 652         return 0;
 653 }
 654
 655
 656 /*
 657  * NAME:        lmGroupCommit()
 658  *
 659  * FUNCTION:    group commit
 660  *      initiate pageout of the pages with COMMIT in the order of
 661  *      page number - redrive pageout of the page at the head of
 662  *      pageout queue until full page has been written.
 663  *
 664  * RETURN:
 665  *
 666  * NOTE:
 667  *      LOGGC_LOCK serializes log group commit queue, and
 668  *      transaction blocks on the commit queue.
 669  *      N.B. LOG_LOCK is NOT held during lmGroupCommit().
 670  */
 671 int lmGroupCommit(struct jfs_log * log, struct tblock * tblk)
 672 {
 673         int rc = 0;
 674
 675         LOGGC_LOCK(log);
 676
 677         /* group committed already ? */
 678         if (tblk->flag & tblkGC_COMMITTED) {
 679                 if (tblk->flag & tblkGC_ERROR)
 680                         rc = -EIO;
 681
 682                 LOGGC_UNLOCK(log);
 683                 return rc;
 684         }
 685         jfs_info("lmGroup Commit: tblk = 0x%p, gcrtc = %d", tblk, log->gcrtc);
 686
 687         if (tblk->xflag & COMMIT_LAZY)
 688                 tblk->flag |= tblkGC_LAZY;
 689
 690         if ((!(log->cflag & logGC_PAGEOUT)) && (!list_empty(&log->cqueue)) &&
 691             (!(tblk->xflag & COMMIT_LAZY) || test_bit(log_FLUSH, &log->flag)
 692              || jfs_tlocks_low)) {
 693                 /*
 694                  * No pageout in progress
 695                  *
 696                  * start group commit as its group leader.
 697                  */
 698                 log->cflag |= logGC_PAGEOUT;
 699
 700                 lmGCwrite(log, 0);
 701         }
 702
 703         if (tblk->xflag & COMMIT_LAZY) {
 704                 /*
 705                  * Lazy transactions can leave now
 706                  */
 707                 LOGGC_UNLOCK(log);
 708                 return 0;
 709         }
 710
 711         /* lmGCwrite gives up LOGGC_LOCK, check again */
 712
 713         if (tblk->flag & tblkGC_COMMITTED) {
 714                 if (tblk->flag & tblkGC_ERROR)
 715                         rc = -EIO;
 716
 717                 LOGGC_UNLOCK(log);
 718                 return rc;
 719         }
 720
 721         /* upcount transaction waiting for completion
 722          */
 723         log->gcrtc++;
 724         tblk->flag |= tblkGC_READY;
 725
 726         __SLEEP_COND(tblk->gcwait, (tblk->flag & tblkGC_COMMITTED),
 727                      LOGGC_LOCK(log), LOGGC_UNLOCK(log));
 728
 729         /* removed from commit queue */
 730         if (tblk->flag & tblkGC_ERROR)
 731                 rc = -EIO;
 732
 733         LOGGC_UNLOCK(log);
 734         return rc;
 735 }
 736
 737 /*
 738  * NAME:        lmGCwrite()
 739  *
 740  * FUNCTION:    group commit write
 741  *      initiate write of log page, building a group of all transactions
 742  *      with commit records on that page.
 743  *
 744  * RETURN:      None
 745  *
 746  * NOTE:
 747  *      LOGGC_LOCK must be held by caller.
 748  *      N.B. LOG_LOCK is NOT held during lmGroupCommit().
 749  */
 750 static void lmGCwrite(struct jfs_log * log, int cant_write)
 751 {
 752         struct lbuf *bp;
 753         struct logpage *lp;
 754         int gcpn;               /* group commit page number */
 755         struct tblock *tblk;
 756         struct tblock *xtblk = NULL;
 757
 758         /*
 759          * build the commit group of a log page
 760          *
 761          * scan commit queue and make a commit group of all
 762          * transactions with COMMIT records on the same log page.
 763          */
 764         /* get the head tblk on the commit queue */
 765         gcpn = list_entry(log->cqueue.next, struct tblock, cqueue)->pn;
 766
 767         list_for_each_entry(tblk, &log->cqueue, cqueue) {
 768                 if (tblk->pn != gcpn)
 769                         break;
 770
 771                 xtblk = tblk;
 772
 773                 /* state transition: (QUEUE, READY) -> COMMIT */
 774                 tblk->flag |= tblkGC_COMMIT;
 775         }
 776         tblk = xtblk;           /* last tblk of the page */
 777
 778         /*
 779          * pageout to commit transactions on the log page.
 780          */
 781         bp = (struct lbuf *) tblk->bp;
 782         lp = (struct logpage *) bp->l_ldata;
 783         /* is page already full ? */
 784         if (tblk->flag & tblkGC_EOP) {
 785                 /* mark page to free at end of group commit of the page */
 786                 tblk->flag &= ~tblkGC_EOP;
 787                 tblk->flag |= tblkGC_FREE;
 788                 bp->l_ceor = bp->l_eor;
 789                 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
 790                 lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmGC,
 791                          cant_write);
 792                 INCREMENT(lmStat.full_page);
 793         }
 794         /* page is not yet full */
 795         else {
 796                 bp->l_ceor = tblk->eor; /* ? bp->l_ceor = bp->l_eor; */
 797                 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
 798                 lbmWrite(log, bp, lbmWRITE | lbmGC, cant_write);
 799                 INCREMENT(lmStat.partial_page);
 800         }
 801 }
 802
 803 /*
 804  * NAME:        lmPostGC()
 805  *
 806  * FUNCTION:    group commit post-processing
 807  *      Processes transactions after their commit records have been written
 808  *      to disk, redriving log I/O if necessary.
 809  *
 810  * RETURN:      None
 811  *
 812  * NOTE:
 813  *      This routine is called a interrupt time by lbmIODone
 814  */
 815 static void lmPostGC(struct lbuf * bp)
 816 {
 817         unsigned long flags;
 818         struct jfs_log *log = bp->l_log;
 819         struct logpage *lp;
 820         struct tblock *tblk, *temp;
 821
 822         //LOGGC_LOCK(log);
 823         spin_lock_irqsave(&log->gclock, flags);
 824         /*
 825          * current pageout of group commit completed.
 826          *
 827          * remove/wakeup transactions from commit queue who were
 828          * group committed with the current log page
 829          */
 830         list_for_each_entry_safe(tblk, temp, &log->cqueue, cqueue) {
 831                 if (!(tblk->flag & tblkGC_COMMIT))
 832                         break;
 833                 /* if transaction was marked GC_COMMIT then
 834                  * it has been shipped in the current pageout
 835                  * and made it to disk - it is committed.
 836                  */
 837
 838                 if (bp->l_flag & lbmERROR)
 839                         tblk->flag |= tblkGC_ERROR;
 840
 841                 /* remove it from the commit queue */
 842                 list_del(&tblk->cqueue);
 843                 tblk->flag &= ~tblkGC_QUEUE;
 844
 845                 if (tblk == log->flush_tblk) {
 846                         /* we can stop flushing the log now */
 847                         clear_bit(log_FLUSH, &log->flag);
 848                         log->flush_tblk = NULL;
 849                 }
 850
 851                 jfs_info("lmPostGC: tblk = 0x%p, flag = 0x%x", tblk,
 852                          tblk->flag);
 853
 854                 if (!(tblk->xflag & COMMIT_FORCE))
 855                         /*
 856                          * Hand tblk over to lazy commit thread
 857                          */
 858                         txLazyUnlock(tblk);
 859                 else {
 860                         /* state transition: COMMIT -> COMMITTED */
 861                         tblk->flag |= tblkGC_COMMITTED;
 862
 863                         if (tblk->flag & tblkGC_READY)
 864                                 log->gcrtc--;
 865
 866                         LOGGC_WAKEUP(tblk);
 867                 }
 868
 869                 /* was page full before pageout ?
 870                  * (and this is the last tblk bound with the page)
 871                  */
 872                 if (tblk->flag & tblkGC_FREE)
 873                         lbmFree(bp);
 874                 /* did page become full after pageout ?
 875                  * (and this is the last tblk bound with the page)
 876                  */
 877                 else if (tblk->flag & tblkGC_EOP) {
 878                         /* finalize the page */
 879                         lp = (struct logpage *) bp->l_ldata;
 880                         bp->l_ceor = bp->l_eor;
 881                         lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
 882                         jfs_info("lmPostGC: calling lbmWrite");
 883                         lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmFREE,
 884                                  1);
 885                 }
 886
 887         }
 888
 889         /* are there any transactions who have entered lnGroupCommit()
 890          * (whose COMMITs are after that of the last log page written.
 891          * They are waiting for new group commit (above at (SLEEP 1))
 892          * or lazy transactions are on a full (queued) log page,
 893          * select the latest ready transaction as new group leader and
 894          * wake her up to lead her group.
 895          */
 896         if ((!list_empty(&log->cqueue)) &&
 897             ((log->gcrtc > 0) || (tblk->bp->l_wqnext != NULL) ||
 898              test_bit(log_FLUSH, &log->flag) || jfs_tlocks_low))
 899                 /*
 900                  * Call lmGCwrite with new group leader
 901                  */
 902                 lmGCwrite(log, 1);
 903
 904         /* no transaction are ready yet (transactions are only just
 905          * queued (GC_QUEUE) and not entered for group commit yet).
 906          * the first transaction entering group commit
 907          * will elect herself as new group leader.
 908          */
 909         else
 910                 log->cflag &= ~logGC_PAGEOUT;
 911
 912         //LOGGC_UNLOCK(log);
 913         spin_unlock_irqrestore(&log->gclock, flags);
 914         return;
 915 }
 916
 917 /*
 918  * NAME:        lmLogSync()
 919  *
 920  * FUNCTION:    write log SYNCPT record for specified log
 921  *      if new sync address is available
 922  *      (normally the case if sync() is executed by back-ground
 923  *      process).
 924  *      if not, explicitly run jfs_blogsync() to initiate
 925  *      getting of new sync address.
 926  *      calculate new value of i_nextsync which determines when
 927  *      this code is called again.
 928  *
 929  *      this is called only from lmLog().
 930  *
 931  * PARAMETER:   ip      - pointer to logs inode.
 932  *
 933  * RETURN:      0
 934  *
 935  * serialization: LOG_LOCK() held on entry/exit
 936  */
 937 static int lmLogSync(struct jfs_log * log, int nosyncwait)
 938 {
 939         int logsize;
 940         int written;            /* written since last syncpt */
 941         int free;               /* free space left available */
 942         int delta;              /* additional delta to write normally */
 943         int more;               /* additional write granted */
 944         struct lrd lrd;
 945         int lsn;
 946         struct logsyncblk *lp;
 947
 948         /*
 949          *      forward syncpt
 950          */
 951         /* if last sync is same as last syncpt,
 952          * invoke sync point forward processing to update sync.
 953          */
 954
 955         if (log->sync == log->syncpt) {
 956                 LOGSYNC_LOCK(log);
 957                 /* ToDo: push dirty metapages out to disk */
 958 //              bmLogSync(log);
 959
 960                 if (list_empty(&log->synclist))
 961                         log->sync = log->lsn;
 962                 else {
 963                         lp = list_entry(log->synclist.next,
 964                                         struct logsyncblk, synclist);
 965                         log->sync = lp->lsn;
 966                 }
 967                 LOGSYNC_UNLOCK(log);
 968
 969         }
 970
 971         /* if sync is different from last syncpt,
 972          * write a SYNCPT record with syncpt = sync.
 973          * reset syncpt = sync
 974          */
 975         if (log->sync != log->syncpt) {
 976                 struct jfs_sb_info *sbi;
 977
 978                 /*
 979                  * We need to make sure all of the "written" metapages
 980                  * actually make it to disk
 981                  */
 982                 list_for_each_entry(sbi, &log->sb_list, log_list) {
 983                         filemap_fdatawrite(sbi->ipbmap->i_mapping);
 984                         filemap_fdatawrite(sbi->ipimap->i_mapping);
 985                         filemap_fdatawrite(sbi->sb->s_bdev->bd_inode->i_mapping);
 986                 }
 987                 list_for_each_entry(sbi, &log->sb_list, log_list) {
 988                         filemap_fdatawait(sbi->ipbmap->i_mapping);
 989                         filemap_fdatawait(sbi->ipimap->i_mapping);
 990                         filemap_fdatawait(sbi->sb->s_bdev->bd_inode->i_mapping);
 991                 }
 992
 993                 lrd.logtid = 0;
 994                 lrd.backchain = 0;
 995                 lrd.type = cpu_to_le16(LOG_SYNCPT);
 996                 lrd.length = 0;
 997                 lrd.log.syncpt.sync = cpu_to_le32(log->sync);
 998                 lsn = lmWriteRecord(log, NULL, &lrd, NULL);
 999
1000                 log->syncpt = log->sync;
1001         } else
1002                 lsn = log->lsn;
1003
1004         /*
1005          *      setup next syncpt trigger (SWAG)
1006          */
1007         logsize = log->logsize;
1008
1009         logdiff(written, lsn, log);
1010         free = logsize - written;
1011         delta = LOGSYNC_DELTA(logsize);
1012         more = min(free / 2, delta);
1013         if (more < 2 * LOGPSIZE) {
1014                 jfs_warn("\n ... Log Wrap ... Log Wrap ... Log Wrap ...\n");
1015                 /*
1016                  *      log wrapping
1017                  *
1018                  * option 1 - panic ? No.!
1019                  * option 2 - shutdown file systems
1020                  *            associated with log ?
1021                  * option 3 - extend log ?
1022                  */
1023                 /*
1024                  * option 4 - second chance
1025                  *
1026                  * mark log wrapped, and continue.
1027                  * when all active transactions are completed,
1028                  * mark log vaild for recovery.
1029                  * if crashed during invalid state, log state
1030                  * implies invald log, forcing fsck().
1031                  */
1032                 /* mark log state log wrap in log superblock */
1033                 /* log->state = LOGWRAP; */
1034
1035                 /* reset sync point computation */
1036                 log->syncpt = log->sync = lsn;
1037                 log->nextsync = delta;
1038         } else
1039                 /* next syncpt trigger = written + more */
1040                 log->nextsync = written + more;
1041
1042         /* return if lmLogSync() from outside of transaction, e.g., sync() */
1043         if (nosyncwait)
1044                 return lsn;
1045
1046         /* if number of bytes written from last sync point is more
1047          * than 1/4 of the log size, stop new transactions from
1048          * starting until all current transactions are completed
1049          * by setting syncbarrier flag.
1050          */
1051         if (written > LOGSYNC_BARRIER(logsize) && logsize > 32 * LOGPSIZE) {
1052                 set_bit(log_SYNCBARRIER, &log->flag);
1053                 jfs_info("log barrier on: lsn=0x%x syncpt=0x%x", lsn,
1054                          log->syncpt);
1055                 /*
1056                  * We may have to initiate group commit
1057                  */
1058                 jfs_flush_journal(log, 0);
1059         }
1060
1061         return lsn;
1062 }
1063
1064
1065 /*
1066  * NAME:        lmLogOpen()
1067  *
1068  * FUNCTION:    open the log on first open;
1069  *      insert filesystem in the active list of the log.
1070  *
1071  * PARAMETER:   ipmnt   - file system mount inode
1072  *              iplog   - log inode (out)
1073  *
1074  * RETURN:
1075  *
1076  * serialization:
1077  */
1078 int lmLogOpen(struct super_block *sb)
1079 {
1080         int rc;
1081         struct block_device *bdev;
1082         struct jfs_log *log;
1083         struct jfs_sb_info *sbi = JFS_SBI(sb);
1084
1085         if (sbi->flag & JFS_NOINTEGRITY)
1086                 return open_dummy_log(sb);
1087
1088         if (sbi->mntflag & JFS_INLINELOG)
1089                 return open_inline_log(sb);
1090
1091         down(&jfs_log_sem);
1092         list_for_each_entry(log, &jfs_external_logs, journal_list) {
1093                 if (log->bdev->bd_dev == sbi->logdev) {
1094                         if (memcmp(log->uuid, sbi->loguuid,
1095                                    sizeof(log->uuid))) {
1096                                 jfs_warn("wrong uuid on JFS journal\n");
1097                                 up(&jfs_log_sem);
1098                                 return -EINVAL;
1099                         }
1100                         /*
1101                          * add file system to log active file system list
1102                          */
1103                         if ((rc = lmLogFileSystem(log, sbi, 1))) {
1104                                 up(&jfs_log_sem);
1105                                 return rc;
1106                         }
1107                         goto journal_found;
1108                 }
1109         }
1110
1111         if (!(log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL))) {
1112                 up(&jfs_log_sem);
1113                 return -ENOMEM;
1114         }
1115         memset(log, 0, sizeof(struct jfs_log));
1116         INIT_LIST_HEAD(&log->sb_list);
1117
1118         /*
1119          *      external log as separate logical volume
1120          *
1121          * file systems to log may have n-to-1 relationship;
1122          */
1123
1124         bdev = open_by_devnum(sbi->logdev, FMODE_READ|FMODE_WRITE);
1125         if (IS_ERR(bdev)) {
1126                 rc = -PTR_ERR(bdev);
1127                 goto free;
1128         }
1129
1130         if ((rc = bd_claim(bdev, log))) {
1131                 goto close;
1132         }
1133
1134         log->bdev = bdev;
1135         memcpy(log->uuid, sbi->loguuid, sizeof(log->uuid));
1136
1137         /*
1138          * initialize log:
1139          */
1140         if ((rc = lmLogInit(log)))
1141                 goto unclaim;
1142
1143         list_add(&log->journal_list, &jfs_external_logs);
1144
1145         /*
1146          * add file system to log active file system list
1147          */
1148         if ((rc = lmLogFileSystem(log, sbi, 1)))
1149                 goto shutdown;
1150
1151 journal_found:
1152         LOG_LOCK(log);
1153         list_add(&sbi->log_list, &log->sb_list);
1154         sbi->log = log;
1155         LOG_UNLOCK(log);
1156
1157         up(&jfs_log_sem);
1158         return 0;
1159
1160         /*
1161          *      unwind on error
1162          */
1163       shutdown:         /* unwind lbmLogInit() */
1164         list_del(&log->journal_list);
1165         lbmLogShutdown(log);
1166
1167       unclaim:
1168         bd_release(bdev);
1169
1170       close:            /* close external log device */
1171         blkdev_put(bdev);
1172
1173       free:             /* free log descriptor */
1174         up(&jfs_log_sem);
1175         kfree(log);
1176
1177         jfs_warn("lmLogOpen: exit(%d)", rc);
1178         return rc;
1179 }
1180
1181 static int open_inline_log(struct super_block *sb)
1182 {
1183         struct jfs_log *log;
1184         int rc;
1185
1186         if (!(log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL)))
1187                 return -ENOMEM;
1188         memset(log, 0, sizeof(struct jfs_log));
1189         INIT_LIST_HEAD(&log->sb_list);
1190
1191         set_bit(log_INLINELOG, &log->flag);
1192         log->bdev = sb->s_bdev;
1193         log->base = addressPXD(&JFS_SBI(sb)->logpxd);
1194         log->size = lengthPXD(&JFS_SBI(sb)->logpxd) >>
1195             (L2LOGPSIZE - sb->s_blocksize_bits);
1196         log->l2bsize = sb->s_blocksize_bits;
1197         ASSERT(L2LOGPSIZE >= sb->s_blocksize_bits);
1198
1199         /*
1200          * initialize log.
1201          */
1202         if ((rc = lmLogInit(log))) {
1203                 kfree(log);
1204                 jfs_warn("lmLogOpen: exit(%d)", rc);
1205                 return rc;
1206         }
1207
1208         list_add(&JFS_SBI(sb)->log_list, &log->sb_list);
1209         JFS_SBI(sb)->log = log;
1210
1211         return rc;
1212 }
1213
1214 static int open_dummy_log(struct super_block *sb)
1215 {
1216         int rc;
1217
1218         down(&jfs_log_sem);
1219         if (!dummy_log) {
1220                 dummy_log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL);
1221                 if (!dummy_log) {
1222                         up(&jfs_log_sem);
1223                         return -ENOMEM;
1224                 }
1225                 memset(dummy_log, 0, sizeof(struct jfs_log));
1226                 INIT_LIST_HEAD(&dummy_log->sb_list);
1227                 dummy_log->no_integrity = 1;
1228                 /* Make up some stuff */
1229                 dummy_log->base = 0;
1230                 dummy_log->size = 1024;
1231                 rc = lmLogInit(dummy_log);
1232                 if (rc) {
1233                         kfree(dummy_log);
1234                         dummy_log = NULL;
1235                         up(&jfs_log_sem);
1236                         return rc;
1237                 }
1238         }
1239
1240         LOG_LOCK(dummy_log);
1241         list_add(&JFS_SBI(sb)->log_list, &dummy_log->sb_list);
1242         JFS_SBI(sb)->log = dummy_log;
1243         LOG_UNLOCK(dummy_log);
1244         up(&jfs_log_sem);
1245
1246         return 0;
1247 }
1248
1249 /*
1250  * NAME:        lmLogInit()
1251  *
1252  * FUNCTION:    log initialization at first log open.
1253  *
1254  *      logredo() (or logformat()) should have been run previously.
1255  *      initialize the log from log superblock.
1256  *      set the log state in the superblock to LOGMOUNT and
1257  *      write SYNCPT log record.
1258  *
1259  * PARAMETER:   log     - log structure
1260  *
1261  * RETURN:      0       - if ok
1262  *              -EINVAL - bad log magic number or superblock dirty
1263  *              error returned from logwait()
1264  *
1265  * serialization: single first open thread
1266  */
1267 int lmLogInit(struct jfs_log * log)
1268 {
1269         int rc = 0;
1270         struct lrd lrd;
1271         struct logsuper *logsuper;
1272         struct lbuf *bpsuper;
1273         struct lbuf *bp;
1274         struct logpage *lp;
1275         int lsn = 0;
1276
1277         jfs_info("lmLogInit: log:0x%p", log);
1278
1279         /* initialize the group commit serialization lock */
1280         LOGGC_LOCK_INIT(log);
1281
1282         /* allocate/initialize the log write serialization lock */
1283         LOG_LOCK_INIT(log);
1284
1285         LOGSYNC_LOCK_INIT(log);
1286
1287         INIT_LIST_HEAD(&log->synclist);
1288
1289         init_waitqueue_head(&log->syncwait);
1290
1291         INIT_LIST_HEAD(&log->cqueue);
1292         log->flush_tblk = NULL;
1293
1294         log->count = 0;
1295
1296         /*
1297          * initialize log i/o
1298          */
1299         if ((rc = lbmLogInit(log)))
1300                 return rc;
1301
1302         if (!test_bit(log_INLINELOG, &log->flag))
1303                 log->l2bsize = L2LOGPSIZE;
1304
1305         /* check for disabled journaling to disk */
1306         if (log->no_integrity) {
1307                 /*
1308                  * Journal pages will still be filled.  When the time comes
1309                  * to actually do the I/O, the write is not done, and the
1310                  * endio routine is called directly.
1311                  */
1312                 bp = lbmAllocate(log , 0);
1313                 log->bp = bp;
1314                 bp->l_pn = bp->l_eor = 0;
1315         } else {
1316                 /*
1317                  * validate log superblock
1318                  */
1319                 if ((rc = lbmRead(log, 1, &bpsuper)))
1320                         goto errout10;
1321
1322                 logsuper = (struct logsuper *) bpsuper->l_ldata;
1323
1324                 if (logsuper->magic != cpu_to_le32(LOGMAGIC)) {
1325                         jfs_warn("*** Log Format Error ! ***");
1326                         rc = -EINVAL;
1327                         goto errout20;
1328                 }
1329
1330                 /* logredo() should have been run successfully. */
1331                 if (logsuper->state != cpu_to_le32(LOGREDONE)) {
1332                         jfs_warn("*** Log Is Dirty ! ***");
1333                         rc = -EINVAL;
1334                         goto errout20;
1335                 }
1336
1337                 /* initialize log from log superblock */
1338                 if (test_bit(log_INLINELOG,&log->flag)) {
1339                         if (log->size != le32_to_cpu(logsuper->size)) {
1340                                 rc = -EINVAL;
1341                                 goto errout20;
1342                         }
1343                         jfs_info("lmLogInit: inline log:0x%p base:0x%Lx "
1344                                  "size:0x%x", log,
1345                                  (unsigned long long) log->base, log->size);
1346                 } else {
1347                         if (memcmp(logsuper->uuid, log->uuid, 16)) {
1348                                 jfs_warn("wrong uuid on JFS log device");
1349                                 goto errout20;
1350                         }
1351                         log->size = le32_to_cpu(logsuper->size);
1352                         log->l2bsize = le32_to_cpu(logsuper->l2bsize);
1353                         jfs_info("lmLogInit: external log:0x%p base:0x%Lx "
1354                                  "size:0x%x", log,
1355                                  (unsigned long long) log->base, log->size);
1356                 }
1357
1358                 log->page = le32_to_cpu(logsuper->end) / LOGPSIZE;
1359                 log->eor = le32_to_cpu(logsuper->end) - (LOGPSIZE * log->page);
1360
1361                 /*
1362                  * initialize for log append write mode
1363                  */
1364                 /* establish current/end-of-log page/buffer */
1365                 if ((rc = lbmRead(log, log->page, &bp)))
1366                         goto errout20;
1367
1368                 lp = (struct logpage *) bp->l_ldata;
1369
1370                 jfs_info("lmLogInit: lsn:0x%x page:%d eor:%d:%d",
1371                          le32_to_cpu(logsuper->end), log->page, log->eor,
1372                          le16_to_cpu(lp->h.eor));
1373
1374                 log->bp = bp;
1375                 bp->l_pn = log->page;
1376                 bp->l_eor = log->eor;
1377
1378                 /* if current page is full, move on to next page */
1379                 if (log->eor >= LOGPSIZE - LOGPTLRSIZE)
1380                         lmNextPage(log);
1381
1382                 /*
1383                  * initialize log syncpoint
1384                  */
1385                 /*
1386                  * write the first SYNCPT record with syncpoint = 0
1387                  * (i.e., log redo up to HERE !);
1388                  * remove current page from lbm write queue at end of pageout
1389                  * (to write log superblock update), but do not release to
1390                  * freelist;
1391                  */
1392                 lrd.logtid = 0;
1393                 lrd.backchain = 0;
1394                 lrd.type = cpu_to_le16(LOG_SYNCPT);
1395                 lrd.length = 0;
1396                 lrd.log.syncpt.sync = 0;
1397                 lsn = lmWriteRecord(log, NULL, &lrd, NULL);
1398                 bp = log->bp;
1399                 bp->l_ceor = bp->l_eor;
1400                 lp = (struct logpage *) bp->l_ldata;
1401                 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
1402                 lbmWrite(log, bp, lbmWRITE | lbmSYNC, 0);
1403                 if ((rc = lbmIOWait(bp, 0)))
1404                         goto errout30;
1405
1406                 /*
1407                  * update/write superblock
1408                  */
1409                 logsuper->state = cpu_to_le32(LOGMOUNT);
1410                 log->serial = le32_to_cpu(logsuper->serial) + 1;
1411                 logsuper->serial = cpu_to_le32(log->serial);
1412                 lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
1413                 if ((rc = lbmIOWait(bpsuper, lbmFREE)))
1414                         goto errout30;
1415         }
1416
1417         /* initialize logsync parameters */
1418         log->logsize = (log->size - 2) << L2LOGPSIZE;
1419         log->lsn = lsn;
1420         log->syncpt = lsn;
1421         log->sync = log->syncpt;
1422         log->nextsync = LOGSYNC_DELTA(log->logsize);
1423
1424         jfs_info("lmLogInit: lsn:0x%x syncpt:0x%x sync:0x%x",
1425                  log->lsn, log->syncpt, log->sync);
1426
1427         /*
1428          * initialize for lazy/group commit
1429          */
1430         log->clsn = lsn;
1431
1432         return 0;
1433
1434         /*
1435          *      unwind on error
1436          */
1437       errout30:         /* release log page */
1438         log->wqueue = NULL;
1439         bp->l_wqnext = NULL;
1440         lbmFree(bp);
1441
1442       errout20:         /* release log superblock */
1443         lbmFree(bpsuper);
1444
1445       errout10:         /* unwind lbmLogInit() */
1446         lbmLogShutdown(log);
1447
1448         jfs_warn("lmLogInit: exit(%d)", rc);
1449         return rc;
1450 }
1451
1452
1453 /*
1454  * NAME:        lmLogClose()
1455  *
1456  * FUNCTION:    remove file system <ipmnt> from active list of log <iplog>
1457  *              and close it on last close.
1458  *
1459  * PARAMETER:   sb      - superblock
1460  *
1461  * RETURN:      errors from subroutines
1462  *
1463  * serialization:
1464  */
1465 int lmLogClose(struct super_block *sb)
1466 {
1467         struct jfs_sb_info *sbi = JFS_SBI(sb);
1468         struct jfs_log *log = sbi->log;
1469         struct block_device *bdev;
1470         int rc = 0;
1471
1472         jfs_info("lmLogClose: log:0x%p", log);
1473
1474         down(&jfs_log_sem);
1475         LOG_LOCK(log);
1476         list_del(&sbi->log_list);
1477         LOG_UNLOCK(log);
1478         sbi->log = NULL;
1479
1480         /*
1481          * We need to make sure all of the "written" metapages
1482          * actually make it to disk
1483          */
1484         sync_blockdev(sb->s_bdev);
1485
1486         if (test_bit(log_INLINELOG, &log->flag)) {
1487                 /*
1488                  *      in-line log in host file system
1489                  */
1490                 rc = lmLogShutdown(log);
1491                 kfree(log);
1492                 goto out;
1493         }
1494
1495         if (!log->no_integrity)
1496                 lmLogFileSystem(log, sbi, 0);
1497
1498         if (!list_empty(&log->sb_list))
1499                 goto out;
1500
1501         /*
1502          * TODO: ensure that the dummy_log is in a state to allow
1503          * lbmLogShutdown to deallocate all the buffers and call
1504          * kfree against dummy_log.  For now, leave dummy_log & its
1505          * buffers in memory, and resuse if another no-integrity mount
1506          * is requested.
1507          */
1508         if (log->no_integrity)
1509                 goto out;
1510
1511         /*
1512          *      external log as separate logical volume
1513          */
1514         list_del(&log->journal_list);
1515         bdev = log->bdev;
1516         rc = lmLogShutdown(log);
1517
1518         bd_release(bdev);
1519         blkdev_put(bdev);
1520
1521         kfree(log);
1522
1523       out:
1524         up(&jfs_log_sem);
1525         jfs_info("lmLogClose: exit(%d)", rc);
1526         return rc;
1527 }
1528
1529
1530 /*
1531  * NAME:        jfs_flush_journal()
1532  *
1533  * FUNCTION:    initiate write of any outstanding transactions to the journal
1534  *              and optionally wait until they are all written to disk
1535  *
1536  *              wait == 0  flush until latest txn is committed, don't wait
1537  *              wait == 1  flush until latest txn is committed, wait
1538  *              wait > 1   flush until all txn's are complete, wait
1539  */
1540 void jfs_flush_journal(struct jfs_log *log, int wait)
1541 {
1542         int i;
1543         struct tblock *target = NULL;
1544
1545         /* jfs_write_inode may call us during read-only mount */
1546         if (!log)
1547                 return;
1548
1549         jfs_info("jfs_flush_journal: log:0x%p wait=%d", log, wait);
1550
1551         LOGGC_LOCK(log);
1552
1553         if (!list_empty(&log->cqueue)) {
1554                 /*
1555                  * This ensures that we will keep writing to the journal as long
1556                  * as there are unwritten commit records
1557                  */
1558                 target = list_entry(log->cqueue.prev, struct tblock, cqueue);
1559
1560                 if (test_bit(log_FLUSH, &log->flag)) {
1561                         /*
1562                          * We're already flushing.
1563                          * if flush_tblk is NULL, we are flushing everything,
1564                          * so leave it that way.  Otherwise, update it to the
1565                          * latest transaction
1566                          */
1567                         if (log->flush_tblk)
1568                                 log->flush_tblk = target;
1569                 } else {
1570                         /* Only flush until latest transaction is committed */
1571                         log->flush_tblk = target;
1572                         set_bit(log_FLUSH, &log->flag);
1573
1574                         /*
1575                          * Initiate I/O on outstanding transactions
1576                          */
1577                         if (!(log->cflag & logGC_PAGEOUT)) {
1578                                 log->cflag |= logGC_PAGEOUT;
1579                                 lmGCwrite(log, 0);
1580                         }
1581                 }
1582         }
1583         if ((wait > 1) || test_bit(log_SYNCBARRIER, &log->flag)) {
1584                 /* Flush until all activity complete */
1585                 set_bit(log_FLUSH, &log->flag);
1586                 log->flush_tblk = NULL;
1587         }
1588
1589         if (wait && target && !(target->flag & tblkGC_COMMITTED)) {
1590                 DECLARE_WAITQUEUE(__wait, current);
1591
1592                 add_wait_queue(&target->gcwait, &__wait);
1593                 set_current_state(TASK_UNINTERRUPTIBLE);
1594                 LOGGC_UNLOCK(log);
1595                 schedule();
1596                 current->state = TASK_RUNNING;
1597                 LOGGC_LOCK(log);
1598                 remove_wait_queue(&target->gcwait, &__wait);
1599         }
1600         LOGGC_UNLOCK(log);
1601
1602         if (wait < 2)
1603                 return;
1604
1605         /*
1606          * If there was recent activity, we may need to wait
1607          * for the lazycommit thread to catch up
1608          */
1609         if ((!list_empty(&log->cqueue)) || !list_empty(&log->synclist)) {
1610                 for (i = 0; i < 800; i++) {     /* Too much? */
1611                         current->state = TASK_INTERRUPTIBLE;
1612                         schedule_timeout(HZ / 4);
1613                         if (list_empty(&log->cqueue) &&
1614                             list_empty(&log->synclist))
1615                                 break;
1616                 }
1617         }
1618         assert(list_empty(&log->cqueue));
1619         assert(list_empty(&log->synclist));
1620         clear_bit(log_FLUSH, &log->flag);
1621 }
1622
1623 /*
1624  * NAME:        lmLogShutdown()
1625  *
1626  * FUNCTION:    log shutdown at last LogClose().
1627  *
1628  *              write log syncpt record.
1629  *              update super block to set redone flag to 0.
1630  *
1631  * PARAMETER:   log     - log inode
1632  *
1633  * RETURN:      0       - success
1634  *
1635  * serialization: single last close thread
1636  */
1637 int lmLogShutdown(struct jfs_log * log)
1638 {
1639         int rc;
1640         struct lrd lrd;
1641         int lsn;
1642         struct logsuper *logsuper;
1643         struct lbuf *bpsuper;
1644         struct lbuf *bp;
1645         struct logpage *lp;
1646
1647         jfs_info("lmLogShutdown: log:0x%p", log);
1648
1649         jfs_flush_journal(log, 2);
1650
1651         /*
1652          * write the last SYNCPT record with syncpoint = 0
1653          * (i.e., log redo up to HERE !)
1654          */
1655         lrd.logtid = 0;
1656         lrd.backchain = 0;
1657         lrd.type = cpu_to_le16(LOG_SYNCPT);
1658         lrd.length = 0;
1659         lrd.log.syncpt.sync = 0;
1660
1661         lsn = lmWriteRecord(log, NULL, &lrd, NULL);
1662         bp = log->bp;
1663         lp = (struct logpage *) bp->l_ldata;
1664         lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
1665         lbmWrite(log, log->bp, lbmWRITE | lbmRELEASE | lbmSYNC, 0);
1666         lbmIOWait(log->bp, lbmFREE);
1667
1668         /*
1669          * synchronous update log superblock
1670          * mark log state as shutdown cleanly
1671          * (i.e., Log does not need to be replayed).
1672          */
1673         if ((rc = lbmRead(log, 1, &bpsuper)))
1674                 goto out;
1675
1676         logsuper = (struct logsuper *) bpsuper->l_ldata;
1677         logsuper->state = cpu_to_le32(LOGREDONE);
1678         logsuper->end = cpu_to_le32(lsn);
1679         lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
1680         rc = lbmIOWait(bpsuper, lbmFREE);
1681
1682         jfs_info("lmLogShutdown: lsn:0x%x page:%d eor:%d",
1683                  lsn, log->page, log->eor);
1684
1685       out:
1686         /*
1687          * shutdown per log i/o
1688          */
1689         lbmLogShutdown(log);
1690
1691         if (rc) {
1692                 jfs_warn("lmLogShutdown: exit(%d)", rc);
1693         }
1694         return rc;
1695 }
1696
1697
1698 /*
1699  * NAME:        lmLogFileSystem()
1700  *
1701  * FUNCTION:    insert (<activate> = true)/remove (<activate> = false)
1702  *      file system into/from log active file system list.
1703  *
1704  * PARAMETE:    log     - pointer to logs inode.
1705  *              fsdev   - kdev_t of filesystem.
1706  *              serial  - pointer to returned log serial number
1707  *              activate - insert/remove device from active list.
1708  *
1709  * RETURN:      0       - success
1710  *              errors returned by vms_iowait().
1711  */
1712 static int lmLogFileSystem(struct jfs_log * log, struct jfs_sb_info *sbi,
1713                            int activate)
1714 {
1715         int rc = 0;
1716         int i;
1717         struct logsuper *logsuper;
1718         struct lbuf *bpsuper;
1719         char *uuid = sbi->uuid;
1720
1721         /*
1722          * insert/remove file system device to log active file system list.
1723          */
1724         if ((rc = lbmRead(log, 1, &bpsuper)))
1725                 return rc;
1726
1727         logsuper = (struct logsuper *) bpsuper->l_ldata;
1728         if (activate) {
1729                 for (i = 0; i < MAX_ACTIVE; i++)
1730                         if (!memcmp(logsuper->active[i].uuid, NULL_UUID, 16)) {
1731                                 memcpy(logsuper->active[i].uuid, uuid, 16);
1732                                 sbi->aggregate = i;
1733                                 break;
1734                         }
1735                 if (i == MAX_ACTIVE) {
1736                         jfs_warn("Too many file systems sharing journal!");
1737                         lbmFree(bpsuper);
1738                         return -EMFILE; /* Is there a better rc? */
1739                 }
1740         } else {
1741                 for (i = 0; i < MAX_ACTIVE; i++)
1742                         if (!memcmp(logsuper->active[i].uuid, uuid, 16)) {
1743                                 memcpy(logsuper->active[i].uuid, NULL_UUID, 16);
1744                                 break;
1745                         }
1746                 if (i == MAX_ACTIVE) {
1747                         jfs_warn("Somebody stomped on the journal!");
1748                         lbmFree(bpsuper);
1749                         return -EIO;
1750                 }
1751
1752         }
1753
1754         /*
1755          * synchronous write log superblock:
1756          *
1757          * write sidestream bypassing write queue:
1758          * at file system mount, log super block is updated for
1759          * activation of the file system before any log record
1760          * (MOUNT record) of the file system, and at file system
1761          * unmount, all meta data for the file system has been
1762          * flushed before log super block is updated for deactivation
1763          * of the file system.
1764          */
1765         lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
1766         rc = lbmIOWait(bpsuper, lbmFREE);
1767
1768         return rc;
1769 }
1770
1771 /*
1772  *              log buffer manager (lbm)
1773  *              ------------------------
1774  *
1775  * special purpose buffer manager supporting log i/o requirements.
1776  *
1777  * per log write queue:
1778  * log pageout occurs in serial order by fifo write queue and
1779  * restricting to a single i/o in pregress at any one time.
1780  * a circular singly-linked list
1781  * (log->wrqueue points to the tail, and buffers are linked via
1782  * bp->wrqueue field), and
1783  * maintains log page in pageout ot waiting for pageout in serial pageout.
1784  */
1785
1786 /*
1787  *      lbmLogInit()
1788  *
1789  * initialize per log I/O setup at lmLogInit()
1790  */
1791 static int lbmLogInit(struct jfs_log * log)
1792 {                               /* log inode */
1793         int i;
1794         struct lbuf *lbuf;
1795
1796         jfs_info("lbmLogInit: log:0x%p", log);
1797
1798         /* initialize current buffer cursor */
1799         log->bp = NULL;
1800
1801         /* initialize log device write queue */
1802         log->wqueue = NULL;
1803
1804         /*
1805          * Each log has its own buffer pages allocated to it.  These are
1806          * not managed by the page cache.  This ensures that a transaction
1807          * writing to the log does not block trying to allocate a page from
1808          * the page cache (for the log).  This would be bad, since page
1809          * allocation waits on the kswapd thread that may be committing inodes
1810          * which would cause log activity.  Was that clear?  I'm trying to
1811          * avoid deadlock here.
1812          */
1813         init_waitqueue_head(&log->free_wait);
1814
1815         log->lbuf_free = NULL;
1816
1817         for (i = 0; i < LOGPAGES; i++) {
1818                 lbuf = kmalloc(sizeof(struct lbuf), GFP_KERNEL);
1819                 if (lbuf == 0)
1820                         goto error;
1821                 lbuf->l_ldata = (char *) get_zeroed_page(GFP_KERNEL);
1822                 if (lbuf->l_ldata == 0) {
1823                         kfree(lbuf);
1824                         goto error;
1825                 }
1826                 lbuf->l_log = log;
1827                 init_waitqueue_head(&lbuf->l_ioevent);
1828
1829                 lbuf->l_freelist = log->lbuf_free;
1830                 log->lbuf_free = lbuf;
1831         }
1832
1833         return (0);
1834
1835       error:
1836         lbmLogShutdown(log);
1837         return -ENOMEM;
1838 }
1839
1840
1841 /*
1842  *      lbmLogShutdown()
1843  *
1844  * finalize per log I/O setup at lmLogShutdown()
1845  */
1846 static void lbmLogShutdown(struct jfs_log * log)
1847 {
1848         struct lbuf *lbuf;
1849
1850         jfs_info("lbmLogShutdown: log:0x%p", log);
1851
1852         lbuf = log->lbuf_free;
1853         while (lbuf) {
1854                 struct lbuf *next = lbuf->l_freelist;
1855                 free_page((unsigned long) lbuf->l_ldata);
1856                 kfree(lbuf);
1857                 lbuf = next;
1858         }
1859
1860         log->bp = NULL;
1861 }
1862
1863
1864 /*
1865  *      lbmAllocate()
1866  *
1867  * allocate an empty log buffer
1868  */
1869 static struct lbuf *lbmAllocate(struct jfs_log * log, int pn)
1870 {
1871         struct lbuf *bp;
1872         unsigned long flags;
1873
1874         /*
1875          * recycle from log buffer freelist if any
1876          */
1877         LCACHE_LOCK(flags);
1878         LCACHE_SLEEP_COND(log->free_wait, (bp = log->lbuf_free), flags);
1879         log->lbuf_free = bp->l_freelist;
1880         LCACHE_UNLOCK(flags);
1881
1882         bp->l_flag = 0;
1883
1884         bp->l_wqnext = NULL;
1885         bp->l_freelist = NULL;
1886
1887         bp->l_pn = pn;
1888         bp->l_blkno = log->base + (pn << (L2LOGPSIZE - log->l2bsize));
1889         bp->l_ceor = 0;
1890
1891         return bp;
1892 }
1893
1894
1895 /*
1896  *      lbmFree()
1897  *
1898  * release a log buffer to freelist
1899  */
1900 static void lbmFree(struct lbuf * bp)
1901 {
1902         unsigned long flags;
1903
1904         LCACHE_LOCK(flags);
1905
1906         lbmfree(bp);
1907
1908         LCACHE_UNLOCK(flags);
1909 }
1910
1911 static void lbmfree(struct lbuf * bp)
1912 {
1913         struct jfs_log *log = bp->l_log;
1914
1915         assert(bp->l_wqnext == NULL);
1916
1917         /*
1918          * return the buffer to head of freelist
1919          */
1920         bp->l_freelist = log->lbuf_free;
1921         log->lbuf_free = bp;
1922
1923         wake_up(&log->free_wait);
1924         return;
1925 }
1926
1927
1928 /*
1929  * NAME:        lbmRedrive
1930  *
1931  * FUNCTION:    add a log buffer to the the log redrive list
1932  *
1933  * PARAMETER:
1934  *     bp       - log buffer
1935  *
1936  * NOTES:
1937  *      Takes log_redrive_lock.
1938  */
1939 static inline void lbmRedrive(struct lbuf *bp)
1940 {
1941         unsigned long flags;
1942
1943         spin_lock_irqsave(&log_redrive_lock, flags);
1944         bp->l_redrive_next = log_redrive_list;
1945         log_redrive_list = bp;
1946         spin_unlock_irqrestore(&log_redrive_lock, flags);
1947
1948         wake_up(&jfs_IO_thread_wait);
1949 }
1950
1951
1952 /*
1953  *      lbmRead()
1954  */
1955 static int lbmRead(struct jfs_log * log, int pn, struct lbuf ** bpp)
1956 {
1957         struct bio *bio;
1958         struct lbuf *bp;
1959
1960         /*
1961          * allocate a log buffer
1962          */
1963         *bpp = bp = lbmAllocate(log, pn);
1964         jfs_info("lbmRead: bp:0x%p pn:0x%x", bp, pn);
1965
1966         bp->l_flag |= lbmREAD;
1967
1968         bio = bio_alloc(GFP_NOFS, 1);
1969
1970         bio->bi_sector = bp->l_blkno << (log->l2bsize - 9);
1971         bio->bi_bdev = log->bdev;
1972         bio->bi_io_vec[0].bv_page = virt_to_page(bp->l_ldata);
1973         bio->bi_io_vec[0].bv_len = LOGPSIZE;
1974         bio->bi_io_vec[0].bv_offset = 0;
1975
1976         bio->bi_vcnt = 1;
1977         bio->bi_idx = 0;
1978         bio->bi_size = LOGPSIZE;
1979
1980         bio->bi_end_io = lbmIODone;
1981         bio->bi_private = bp;
1982         submit_bio(READ_SYNC, bio);
1983
1984         wait_event(bp->l_ioevent, (bp->l_flag != lbmREAD));
1985
1986         return 0;
1987 }
1988
1989
1990 /*
1991  *      lbmWrite()
1992  *
1993  * buffer at head of pageout queue stays after completion of
1994  * partial-page pageout and redriven by explicit initiation of
1995  * pageout by caller until full-page pageout is completed and
1996  * released.
1997  *
1998  * device driver i/o done redrives pageout of new buffer at
1999  * head of pageout queue when current buffer at head of pageout
2000  * queue is released at the completion of its full-page pageout.
2001  *
2002  * LOGGC_LOCK() serializes lbmWrite() by lmNextPage() and lmGroupCommit().
2003  * LCACHE_LOCK() serializes xflag between lbmWrite() and lbmIODone()
2004  */
2005 static void lbmWrite(struct jfs_log * log, struct lbuf * bp, int flag,
2006                      int cant_block)
2007 {
2008         struct lbuf *tail;
2009         unsigned long flags;
2010
2011         jfs_info("lbmWrite: bp:0x%p flag:0x%x pn:0x%x", bp, flag, bp->l_pn);
2012
2013         /* map the logical block address to physical block address */
2014         bp->l_blkno =
2015             log->base + (bp->l_pn << (L2LOGPSIZE - log->l2bsize));
2016
2017         LCACHE_LOCK(flags);             /* disable+lock */
2018
2019         /*
2020          * initialize buffer for device driver
2021          */
2022         bp->l_flag = flag;
2023
2024         /*
2025          *      insert bp at tail of write queue associated with log
2026          *
2027          * (request is either for bp already/currently at head of queue
2028          * or new bp to be inserted at tail)
2029          */
2030         tail = log->wqueue;
2031
2032         /* is buffer not already on write queue ? */
2033         if (bp->l_wqnext == NULL) {
2034                 /* insert at tail of wqueue */
2035                 if (tail == NULL) {
2036                         log->wqueue = bp;
2037                         bp->l_wqnext = bp;
2038                 } else {
2039                         log->wqueue = bp;
2040                         bp->l_wqnext = tail->l_wqnext;
2041                         tail->l_wqnext = bp;
2042                 }
2043
2044                 tail = bp;
2045         }
2046
2047         /* is buffer at head of wqueue and for write ? */
2048         if ((bp != tail->l_wqnext) || !(flag & lbmWRITE)) {
2049                 LCACHE_UNLOCK(flags);   /* unlock+enable */
2050                 return;
2051         }
2052
2053         LCACHE_UNLOCK(flags);   /* unlock+enable */
2054
2055         if (cant_block)
2056                 lbmRedrive(bp);
2057         else if (flag & lbmSYNC)
2058                 lbmStartIO(bp);
2059         else {
2060                 LOGGC_UNLOCK(log);
2061                 lbmStartIO(bp);
2062                 LOGGC_LOCK(log);
2063         }
2064 }
2065
2066
2067 /*
2068  *      lbmDirectWrite()
2069  *
2070  * initiate pageout bypassing write queue for sidestream
2071  * (e.g., log superblock) write;
2072  */
2073 static void lbmDirectWrite(struct jfs_log * log, struct lbuf * bp, int flag)
2074 {
2075         jfs_info("lbmDirectWrite: bp:0x%p flag:0x%x pn:0x%x",
2076                  bp, flag, bp->l_pn);
2077
2078         /*
2079          * initialize buffer for device driver
2080          */
2081         bp->l_flag = flag | lbmDIRECT;
2082
2083         /* map the logical block address to physical block address */
2084         bp->l_blkno =
2085             log->base + (bp->l_pn << (L2LOGPSIZE - log->l2bsize));
2086
2087         /*
2088          *      initiate pageout of the page
2089          */
2090         lbmStartIO(bp);
2091 }
2092
2093
2094 /*
2095  * NAME:        lbmStartIO()
2096  *
2097  * FUNCTION:    Interface to DD strategy routine
2098  *
2099  * RETURN:      none
2100  *
2101  * serialization: LCACHE_LOCK() is NOT held during log i/o;
2102  */
2103 static void lbmStartIO(struct lbuf * bp)
2104 {
2105         struct bio *bio;
2106         struct jfs_log *log = bp->l_log;
2107
2108         jfs_info("lbmStartIO\n");
2109
2110         bio = bio_alloc(GFP_NOFS, 1);
2111         bio->bi_sector = bp->l_blkno << (log->l2bsize - 9);
2112         bio->bi_bdev = log->bdev;
2113         bio->bi_io_vec[0].bv_page = virt_to_page(bp->l_ldata);
2114         bio->bi_io_vec[0].bv_len = LOGPSIZE;
2115         bio->bi_io_vec[0].bv_offset = 0;
2116
2117         bio->bi_vcnt = 1;
2118         bio->bi_idx = 0;
2119         bio->bi_size = LOGPSIZE;
2120
2121         bio->bi_end_io = lbmIODone;
2122         bio->bi_private = bp;
2123
2124         /* check if journaling to disk has been disabled */
2125         if (!log->no_integrity) {
2126                 submit_bio(WRITE_SYNC, bio);
2127                 INCREMENT(lmStat.submitted);
2128         }
2129         else {
2130                 bio->bi_size = 0;
2131                 lbmIODone(bio, 0, 0); /* 2nd argument appears to not be used => 0
2132                                        *  3rd argument appears to not be used => 0
2133                                        */
2134         }
2135 }
2136
2137
2138 /*
2139  *      lbmIOWait()
2140  */
2141 static int lbmIOWait(struct lbuf * bp, int flag)
2142 {
2143         unsigned long flags;
2144         int rc = 0;
2145
2146         jfs_info("lbmIOWait1: bp:0x%p flag:0x%x:0x%x", bp, bp->l_flag, flag);
2147
2148         LCACHE_LOCK(flags);             /* disable+lock */
2149
2150         LCACHE_SLEEP_COND(bp->l_ioevent, (bp->l_flag & lbmDONE), flags);
2151
2152         rc = (bp->l_flag & lbmERROR) ? -EIO : 0;
2153
2154         if (flag & lbmFREE)
2155                 lbmfree(bp);
2156
2157         LCACHE_UNLOCK(flags);   /* unlock+enable */
2158
2159         jfs_info("lbmIOWait2: bp:0x%p flag:0x%x:0x%x", bp, bp->l_flag, flag);
2160         return rc;
2161 }
2162
2163 /*
2164  *      lbmIODone()
2165  *
2166  * executed at INTIODONE level
2167  */
2168 static int lbmIODone(struct bio *bio, unsigned int bytes_done, int error)
2169 {
2170         struct lbuf *bp = bio->bi_private;
2171         struct lbuf *nextbp, *tail;
2172         struct jfs_log *log;
2173         unsigned long flags;
2174
2175         if (bio->bi_size)
2176                 return 1;
2177
2178         /*
2179          * get back jfs buffer bound to the i/o buffer
2180          */
2181         jfs_info("lbmIODone: bp:0x%p flag:0x%x", bp, bp->l_flag);
2182
2183         LCACHE_LOCK(flags);             /* disable+lock */
2184
2185         bp->l_flag |= lbmDONE;
2186
2187         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188                 bp->l_flag |= lbmERROR;
2189
2190                 jfs_err("lbmIODone: I/O error in JFS log");
2191         }
2192
2193         bio_put(bio);
2194
2195         /*
2196          *      pagein completion
2197          */
2198         if (bp->l_flag & lbmREAD) {
2199                 bp->l_flag &= ~lbmREAD;
2200
2201                 LCACHE_UNLOCK(flags);   /* unlock+enable */
2202
2203                 /* wakeup I/O initiator */
2204                 LCACHE_WAKEUP(&bp->l_ioevent);
2205
2206                 return 0;
2207         }
2208
2209         /*
2210          *      pageout completion
2211          *
2212          * the bp at the head of write queue has completed pageout.
2213          *
2214          * if single-commit/full-page pageout, remove the current buffer
2215          * from head of pageout queue, and redrive pageout with
2216          * the new buffer at head of pageout queue;
2217          * otherwise, the partial-page pageout buffer stays at
2218          * the head of pageout queue to be redriven for pageout
2219          * by lmGroupCommit() until full-page pageout is completed.
2220          */
2221         bp->l_flag &= ~lbmWRITE;
2222         INCREMENT(lmStat.pagedone);
2223
2224         /* update committed lsn */
2225         log = bp->l_log;
2226         log->clsn = (bp->l_pn << L2LOGPSIZE) + bp->l_ceor;
2227
2228         if (bp->l_flag & lbmDIRECT) {
2229                 LCACHE_WAKEUP(&bp->l_ioevent);
2230                 LCACHE_UNLOCK(flags);
2231                 return 0;
2232         }
2233
2234         tail = log->wqueue;
2235
2236         /* single element queue */
2237         if (bp == tail) {
2238                 /* remove head buffer of full-page pageout
2239                  * from log device write queue
2240                  */
2241                 if (bp->l_flag & lbmRELEASE) {
2242                         log->wqueue = NULL;
2243                         bp->l_wqnext = NULL;
2244                 }
2245         }
2246         /* multi element queue */
2247         else {
2248                 /* remove head buffer of full-page pageout
2249                  * from log device write queue
2250                  */
2251                 if (bp->l_flag & lbmRELEASE) {
2252                         nextbp = tail->l_wqnext = bp->l_wqnext;
2253                         bp->l_wqnext = NULL;
2254
2255                         /*
2256                          * redrive pageout of next page at head of write queue:
2257                          * redrive next page without any bound tblk
2258                          * (i.e., page w/o any COMMIT records), or
2259                          * first page of new group commit which has been
2260                          * queued after current page (subsequent pageout
2261                          * is performed synchronously, except page without
2262                          * any COMMITs) by lmGroupCommit() as indicated
2263                          * by lbmWRITE flag;
2264                          */
2265                         if (nextbp->l_flag & lbmWRITE) {
2266                                 /*
2267                                  * We can't do the I/O at interrupt time.
2268                                  * The jfsIO thread can do it
2269                                  */
2270                                 lbmRedrive(nextbp);
2271                         }
2272                 }
2273         }
2274
2275         /*
2276          *      synchronous pageout:
2277          *
2278          * buffer has not necessarily been removed from write queue
2279          * (e.g., synchronous write of partial-page with COMMIT):
2280          * leave buffer for i/o initiator to dispose
2281          */
2282         if (bp->l_flag & lbmSYNC) {
2283                 LCACHE_UNLOCK(flags);   /* unlock+enable */
2284
2285                 /* wakeup I/O initiator */
2286                 LCACHE_WAKEUP(&bp->l_ioevent);
2287         }
2288
2289         /*
2290          *      Group Commit pageout:
2291          */
2292         else if (bp->l_flag & lbmGC) {
2293                 LCACHE_UNLOCK(flags);
2294                 lmPostGC(bp);
2295         }
2296
2297         /*
2298          *      asynchronous pageout:
2299          *
2300          * buffer must have been removed from write queue:
2301          * insert buffer at head of freelist where it can be recycled
2302          */
2303         else {
2304                 assert(bp->l_flag & lbmRELEASE);
2305                 assert(bp->l_flag & lbmFREE);
2306                 lbmfree(bp);
2307
2308                 LCACHE_UNLOCK(flags);   /* unlock+enable */
2309         }
2310
2311         return 0;
2312 }
2313
2314 int jfsIOWait(void *arg)
2315 {
2316         struct lbuf *bp;
2317
2318         daemonize("jfsIO");
2319
2320         complete(&jfsIOwait);
2321
2322         do {
2323                 DECLARE_WAITQUEUE(wq, current);
2324
2325                 spin_lock_irq(&log_redrive_lock);
2326                 while ((bp = log_redrive_list) != 0) {
2327                         log_redrive_list = bp->l_redrive_next;
2328                         bp->l_redrive_next = NULL;
2329                         spin_unlock_irq(&log_redrive_lock);
2330                         lbmStartIO(bp);
2331                         spin_lock_irq(&log_redrive_lock);
2332                 }
2333                 if (current->flags & PF_FREEZE) {
2334                         spin_unlock_irq(&log_redrive_lock);
2335                         refrigerator(PF_FREEZE);
2336                 } else {
2337                         add_wait_queue(&jfs_IO_thread_wait, &wq);
2338                         set_current_state(TASK_INTERRUPTIBLE);
2339                         spin_unlock_irq(&log_redrive_lock);
2340                         schedule();
2341                         current->state = TASK_RUNNING;
2342                         remove_wait_queue(&jfs_IO_thread_wait, &wq);
2343                 }
2344         } while (!jfs_stop_threads);
2345
2346         jfs_info("jfsIOWait being killed!");
2347         complete_and_exit(&jfsIOwait, 0);
2348 }
2349
2350 /*
2351  * NAME:        lmLogFormat()/jfs_logform()
2352  *
2353  * FUNCTION:    format file system log
2354  *
2355  * PARAMETERS:
2356  *      log     - volume log
2357  *      logAddress - start address of log space in FS block
2358  *      logSize - length of log space in FS block;
2359  *
2360  * RETURN:      0       - success
2361  *              -EIO    - i/o error
2362  *
2363  * XXX: We're synchronously writing one page at a time.  This needs to
2364  *      be improved by writing multiple pages at once.
2365  */
2366 int lmLogFormat(struct jfs_log *log, s64 logAddress, int logSize)
2367 {
2368         int rc = -EIO;
2369         struct jfs_sb_info *sbi;
2370         struct logsuper *logsuper;
2371         struct logpage *lp;
2372         int lspn;               /* log sequence page number */
2373         struct lrd *lrd_ptr;
2374         int npages = 0;
2375         struct lbuf *bp;
2376
2377         jfs_info("lmLogFormat: logAddress:%Ld logSize:%d",
2378                  (long long)logAddress, logSize);
2379
2380         sbi = list_entry(log->sb_list.next, struct jfs_sb_info, log_list);
2381
2382         /* allocate a log buffer */
2383         bp = lbmAllocate(log, 1);
2384
2385         npages = logSize >> sbi->l2nbperpage;
2386
2387         /*
2388          *      log space:
2389          *
2390          * page 0 - reserved;
2391          * page 1 - log superblock;
2392          * page 2 - log data page: A SYNC log record is written
2393          *          into this page at logform time;
2394          * pages 3-N - log data page: set to empty log data pages;
2395          */
2396         /*
2397          *      init log superblock: log page 1
2398          */
2399         logsuper = (struct logsuper *) bp->l_ldata;
2400
2401         logsuper->magic = cpu_to_le32(LOGMAGIC);
2402         logsuper->version = cpu_to_le32(LOGVERSION);
2403         logsuper->state = cpu_to_le32(LOGREDONE);
2404         logsuper->flag = cpu_to_le32(sbi->mntflag);     /* ? */
2405         logsuper->size = cpu_to_le32(npages);
2406         logsuper->bsize = cpu_to_le32(sbi->bsize);
2407         logsuper->l2bsize = cpu_to_le32(sbi->l2bsize);
2408         logsuper->end = cpu_to_le32(2 * LOGPSIZE + LOGPHDRSIZE + LOGRDSIZE);
2409
2410         bp->l_flag = lbmWRITE | lbmSYNC | lbmDIRECT;
2411         bp->l_blkno = logAddress + sbi->nbperpage;
2412         lbmStartIO(bp);
2413         if ((rc = lbmIOWait(bp, 0)))
2414                 goto exit;
2415
2416         /*
2417          *      init pages 2 to npages-1 as log data pages:
2418          *
2419          * log page sequence number (lpsn) initialization:
2420          *
2421          * pn:   0     1     2     3                 n-1
2422          *       +-----+-----+=====+=====+===.....===+=====+
2423          * lspn:             N-1   0     1           N-2
2424          *                   <--- N page circular file ---->
2425          *
2426          * the N (= npages-2) data pages of the log is maintained as
2427          * a circular file for the log records;
2428          * lpsn grows by 1 monotonically as each log page is written
2429          * to the circular file of the log;
2430          * and setLogpage() will not reset the page number even if
2431          * the eor is equal to LOGPHDRSIZE. In order for binary search
2432          * still work in find log end process, we have to simulate the
2433          * log wrap situation at the log format time.
2434          * The 1st log page written will have the highest lpsn. Then
2435          * the succeeding log pages will have ascending order of
2436          * the lspn starting from 0, ... (N-2)
2437          */
2438         lp = (struct logpage *) bp->l_ldata;
2439         /*
2440          * initialize 1st log page to be written: lpsn = N - 1,
2441          * write a SYNCPT log record is written to this page
2442          */
2443         lp->h.page = lp->t.page = cpu_to_le32(npages - 3);
2444         lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE + LOGRDSIZE);
2445
2446         lrd_ptr = (struct lrd *) &lp->data;
2447         lrd_ptr->logtid = 0;
2448         lrd_ptr->backchain = 0;
2449         lrd_ptr->type = cpu_to_le16(LOG_SYNCPT);
2450         lrd_ptr->length = 0;
2451         lrd_ptr->log.syncpt.sync = 0;
2452
2453         bp->l_blkno += sbi->nbperpage;
2454         bp->l_flag = lbmWRITE | lbmSYNC | lbmDIRECT;
2455         lbmStartIO(bp);
2456         if ((rc = lbmIOWait(bp, 0)))
2457                 goto exit;
2458
2459         /*
2460          *      initialize succeeding log pages: lpsn = 0, 1, ..., (N-2)
2461          */
2462         for (lspn = 0; lspn < npages - 3; lspn++) {
2463                 lp->h.page = lp->t.page = cpu_to_le32(lspn);
2464                 lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE);
2465
2466                 bp->l_blkno += sbi->nbperpage;
2467                 bp->l_flag = lbmWRITE | lbmSYNC | lbmDIRECT;
2468                 lbmStartIO(bp);
2469                 if ((rc = lbmIOWait(bp, 0)))
2470                         goto exit;
2471         }
2472
2473         rc = 0;
2474 exit:
2475         /*
2476          *      finalize log
2477          */
2478         /* release the buffer */
2479         lbmFree(bp);
2480
2481         return rc;
2482 }
2483
2484 #ifdef CONFIG_JFS_STATISTICS
2485 int jfs_lmstats_read(char *buffer, char **start, off_t offset, int length,
2486                       int *eof, void *data)
2487 {
2488         int len = 0;
2489         off_t begin;
2490
2491         len += sprintf(buffer,
2492                        "JFS Logmgr stats\n"
2493                        "================\n"
2494                        "commits = %d\n"
2495                        "writes submitted = %d\n"
2496                        "writes completed = %d\n"
2497                        "full pages submitted = %d\n"
2498                        "partial pages submitted = %d\n",
2499                        lmStat.commit,
2500                        lmStat.submitted,
2501                        lmStat.pagedone,
2502                        lmStat.full_page,
2503                        lmStat.partial_page);
2504
2505         begin = offset;
2506         *start = buffer + begin;
2507         len -= begin;
2508
2509         if (len > length)
2510                 len = length;
2511         else
2512                 *eof = 1;
2513
2514         if (len < 0)
2515                 len = 0;
2516
2517         return len;
2518 }
2519 #endif /* CONFIG_JFS_STATISTICS */