fs/xfs/xfs_log_priv.h

   1 /*
   2  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #ifndef __XFS_LOG_PRIV_H__
  19 #define __XFS_LOG_PRIV_H__
  20
  21 struct xfs_buf;
  22 struct log;
  23 struct xlog_ticket;
  24 struct xfs_buf_cancel;
  25 struct xfs_mount;
  26
  27 /*
  28  * Macros, structures, prototypes for internal log manager use.
  29  */
  30
  31 #define XLOG_MIN_ICLOGS         2
  32 #define XLOG_MAX_ICLOGS         8
  33 #define XLOG_HEADER_MAGIC_NUM   0xFEEDbabe      /* Invalid cycle number */
  34 #define XLOG_VERSION_1          1
  35 #define XLOG_VERSION_2          2               /* Large IClogs, Log sunit */
  36 #define XLOG_VERSION_OKBITS     (XLOG_VERSION_1 | XLOG_VERSION_2)
  37 #define XLOG_MIN_RECORD_BSIZE   (16*1024)       /* eventually 32k */
  38 #define XLOG_BIG_RECORD_BSIZE   (32*1024)       /* 32k buffers */
  39 #define XLOG_MAX_RECORD_BSIZE   (256*1024)
  40 #define XLOG_HEADER_CYCLE_SIZE  (32*1024)       /* cycle data in header */
  41 #define XLOG_MIN_RECORD_BSHIFT  14              /* 16384 == 1 << 14 */
  42 #define XLOG_BIG_RECORD_BSHIFT  15              /* 32k == 1 << 15 */
  43 #define XLOG_MAX_RECORD_BSHIFT  18              /* 256k == 1 << 18 */
  44 #define XLOG_BTOLSUNIT(log, b)  (((b)+(log)->l_mp->m_sb.sb_logsunit-1) / \
  45                                  (log)->l_mp->m_sb.sb_logsunit)
  46 #define XLOG_LSUNITTOB(log, su) ((su) * (log)->l_mp->m_sb.sb_logsunit)
  47
  48 #define XLOG_HEADER_SIZE        512
  49
  50 #define XLOG_REC_SHIFT(log) \
  51         BTOBB(1 << (xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? \
  52          XLOG_MAX_RECORD_BSHIFT : XLOG_BIG_RECORD_BSHIFT))
  53 #define XLOG_TOTAL_REC_SHIFT(log) \
  54         BTOBB(XLOG_MAX_ICLOGS << (xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? \
  55          XLOG_MAX_RECORD_BSHIFT : XLOG_BIG_RECORD_BSHIFT))
  56
  57
  58 static inline xfs_lsn_t xlog_assign_lsn(uint cycle, uint block)
  59 {
  60         return ((xfs_lsn_t)cycle << 32) | block;
  61 }
  62
  63 static inline uint xlog_get_cycle(char *ptr)
  64 {
  65         if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
  66                 return be32_to_cpu(*((__be32 *)ptr + 1));
  67         else
  68                 return be32_to_cpu(*(__be32 *)ptr);
  69 }
  70
  71 #define BLK_AVG(blk1, blk2)     ((blk1+blk2) >> 1)
  72
  73 #ifdef __KERNEL__
  74
  75 /*
  76  * get client id from packed copy.
  77  *
  78  * this hack is here because the xlog_pack code copies four bytes
  79  * of xlog_op_header containing the fields oh_clientid, oh_flags
  80  * and oh_res2 into the packed copy.
  81  *
  82  * later on this four byte chunk is treated as an int and the
  83  * client id is pulled out.
  84  *
  85  * this has endian issues, of course.
  86  */
  87 static inline uint xlog_get_client_id(__be32 i)
  88 {
  89         return be32_to_cpu(i) >> 24;
  90 }
  91
  92 #define xlog_panic(args...)     cmn_err(CE_PANIC, ## args)
  93 #define xlog_exit(args...)      cmn_err(CE_PANIC, ## args)
  94 #define xlog_warn(args...)      cmn_err(CE_WARN, ## args)
  95
  96 /*
  97  * In core log state
  98  */
  99 #define XLOG_STATE_ACTIVE    0x0001 /* Current IC log being written to */
 100 #define XLOG_STATE_WANT_SYNC 0x0002 /* Want to sync this iclog; no more writes */
 101 #define XLOG_STATE_SYNCING   0x0004 /* This IC log is syncing */
 102 #define XLOG_STATE_DONE_SYNC 0x0008 /* Done syncing to disk */
 103 #define XLOG_STATE_DO_CALLBACK \
 104                              0x0010 /* Process callback functions */
 105 #define XLOG_STATE_CALLBACK  0x0020 /* Callback functions now */
 106 #define XLOG_STATE_DIRTY     0x0040 /* Dirty IC log, not ready for ACTIVE status*/
 107 #define XLOG_STATE_IOERROR   0x0080 /* IO error happened in sync'ing log */
 108 #define XLOG_STATE_ALL       0x7FFF /* All possible valid flags */
 109 #define XLOG_STATE_NOTUSED   0x8000 /* This IC log not being used */
 110 #endif  /* __KERNEL__ */
 111
 112 /*
 113  * Flags to log operation header
 114  *
 115  * The first write of a new transaction will be preceded with a start
 116  * record, XLOG_START_TRANS.  Once a transaction is committed, a commit
 117  * record is written, XLOG_COMMIT_TRANS.  If a single region can not fit into
 118  * the remainder of the current active in-core log, it is split up into
 119  * multiple regions.  Each partial region will be marked with a
 120  * XLOG_CONTINUE_TRANS until the last one, which gets marked with XLOG_END_TRANS.
 121  *
 122  */
 123 #define XLOG_START_TRANS        0x01    /* Start a new transaction */
 124 #define XLOG_COMMIT_TRANS       0x02    /* Commit this transaction */
 125 #define XLOG_CONTINUE_TRANS     0x04    /* Cont this trans into new region */
 126 #define XLOG_WAS_CONT_TRANS     0x08    /* Cont this trans into new region */
 127 #define XLOG_END_TRANS          0x10    /* End a continued transaction */
 128 #define XLOG_UNMOUNT_TRANS      0x20    /* Unmount a filesystem transaction */
 129
 130 #ifdef __KERNEL__
 131 /*
 132  * Flags to log ticket
 133  */
 134 #define XLOG_TIC_INITED         0x1     /* has been initialized */
 135 #define XLOG_TIC_PERM_RESERV    0x2     /* permanent reservation */
 136 #define XLOG_TIC_IN_Q           0x4
 137
 138 #define XLOG_TIC_FLAGS \
 139         { XLOG_TIC_INITED,      "XLOG_TIC_INITED" }, \
 140         { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }, \
 141         { XLOG_TIC_IN_Q,        "XLOG_TIC_IN_Q" }
 142
 143 #endif  /* __KERNEL__ */
 144
 145 #define XLOG_UNMOUNT_TYPE       0x556e  /* Un for Unmount */
 146
 147 /*
 148  * Flags for log structure
 149  */
 150 #define XLOG_CHKSUM_MISMATCH    0x1     /* used only during recovery */
 151 #define XLOG_ACTIVE_RECOVERY    0x2     /* in the middle of recovery */
 152 #define XLOG_RECOVERY_NEEDED    0x4     /* log was recovered */
 153 #define XLOG_IO_ERROR           0x8     /* log hit an I/O error, and being
 154                                            shutdown */
 155 typedef __uint32_t xlog_tid_t;
 156
 157
 158 #ifdef __KERNEL__
 159 /*
 160  * Below are states for covering allocation transactions.
 161  * By covering, we mean changing the h_tail_lsn in the last on-disk
 162  * log write such that no allocation transactions will be re-done during
 163  * recovery after a system crash. Recovery starts at the last on-disk
 164  * log write.
 165  *
 166  * These states are used to insert dummy log entries to cover
 167  * space allocation transactions which can undo non-transactional changes
 168  * after a crash. Writes to a file with space
 169  * already allocated do not result in any transactions. Allocations
 170  * might include space beyond the EOF. So if we just push the EOF a
 171  * little, the last transaction for the file could contain the wrong
 172  * size. If there is no file system activity, after an allocation
 173  * transaction, and the system crashes, the allocation transaction
 174  * will get replayed and the file will be truncated. This could
 175  * be hours/days/... after the allocation occurred.
 176  *
 177  * The fix for this is to do two dummy transactions when the
 178  * system is idle. We need two dummy transaction because the h_tail_lsn
 179  * in the log record header needs to point beyond the last possible
 180  * non-dummy transaction. The first dummy changes the h_tail_lsn to
 181  * the first transaction before the dummy. The second dummy causes
 182  * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
 183  *
 184  * These dummy transactions get committed when everything
 185  * is idle (after there has been some activity).
 186  *
 187  * There are 5 states used to control this.
 188  *
 189  *  IDLE -- no logging has been done on the file system or
 190  *              we are done covering previous transactions.
 191  *  NEED -- logging has occurred and we need a dummy transaction
 192  *              when the log becomes idle.
 193  *  DONE -- we were in the NEED state and have committed a dummy
 194  *              transaction.
 195  *  NEED2 -- we detected that a dummy transaction has gone to the
 196  *              on disk log with no other transactions.
 197  *  DONE2 -- we committed a dummy transaction when in the NEED2 state.
 198  *
 199  * There are two places where we switch states:
 200  *
 201  * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
 202  *      We commit the dummy transaction and switch to DONE or DONE2,
 203  *      respectively. In all other states, we don't do anything.
 204  *
 205  * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
 206  *
 207  *      No matter what state we are in, if this isn't the dummy
 208  *      transaction going out, the next state is NEED.
 209  *      So, if we aren't in the DONE or DONE2 states, the next state
 210  *      is NEED. We can't be finishing a write of the dummy record
 211  *      unless it was committed and the state switched to DONE or DONE2.
 212  *
 213  *      If we are in the DONE state and this was a write of the
 214  *              dummy transaction, we move to NEED2.
 215  *
 216  *      If we are in the DONE2 state and this was a write of the
 217  *              dummy transaction, we move to IDLE.
 218  *
 219  *
 220  * Writing only one dummy transaction can get appended to
 221  * one file space allocation. When this happens, the log recovery
 222  * code replays the space allocation and a file could be truncated.
 223  * This is why we have the NEED2 and DONE2 states before going idle.
 224  */
 225
 226 #define XLOG_STATE_COVER_IDLE   0
 227 #define XLOG_STATE_COVER_NEED   1
 228 #define XLOG_STATE_COVER_DONE   2
 229 #define XLOG_STATE_COVER_NEED2  3
 230 #define XLOG_STATE_COVER_DONE2  4
 231
 232 #define XLOG_COVER_OPS          5
 233
 234
 235 /* Ticket reservation region accounting */
 236 #define XLOG_TIC_LEN_MAX        15
 237
 238 /*
 239  * Reservation region
 240  * As would be stored in xfs_log_iovec but without the i_addr which
 241  * we don't care about.
 242  */
 243 typedef struct xlog_res {
 244         uint    r_len;  /* region length                :4 */
 245         uint    r_type; /* region's transaction type    :4 */
 246 } xlog_res_t;
 247
 248 typedef struct xlog_ticket {
 249         sv_t               t_wait;       /* ticket wait queue            : 20 */
 250         struct xlog_ticket *t_next;      /*                              :4|8 */
 251         struct xlog_ticket *t_prev;      /*                              :4|8 */
 252         xlog_tid_t         t_tid;        /* transaction identifier       : 4  */
 253         atomic_t           t_ref;        /* ticket reference count       : 4  */
 254         int                t_curr_res;   /* current reservation in bytes : 4  */
 255         int                t_unit_res;   /* unit reservation in bytes    : 4  */
 256         char               t_ocnt;       /* original count               : 1  */
 257         char               t_cnt;        /* current count                : 1  */
 258         char               t_clientid;   /* who does this belong to;     : 1  */
 259         char               t_flags;      /* properties of reservation    : 1  */
 260         uint               t_trans_type; /* transaction type             : 4  */
 261
 262         /* reservation array fields */
 263         uint               t_res_num;                    /* num in array : 4 */
 264         uint               t_res_num_ophdrs;             /* num op hdrs  : 4 */
 265         uint               t_res_arr_sum;                /* array sum    : 4 */
 266         uint               t_res_o_flow;                 /* sum overflow : 4 */
 267         xlog_res_t         t_res_arr[XLOG_TIC_LEN_MAX];  /* array of res : 8 * 15 */
 268 } xlog_ticket_t;
 269
 270 #endif
 271
 272
 273 typedef struct xlog_op_header {
 274         __be32     oh_tid;      /* transaction id of operation  :  4 b */
 275         __be32     oh_len;      /* bytes in data region         :  4 b */
 276         __u8       oh_clientid; /* who sent me this             :  1 b */
 277         __u8       oh_flags;    /*                              :  1 b */
 278         __u16      oh_res2;     /* 32 bit align                 :  2 b */
 279 } xlog_op_header_t;
 280
 281
 282 /* valid values for h_fmt */
 283 #define XLOG_FMT_UNKNOWN  0
 284 #define XLOG_FMT_LINUX_LE 1
 285 #define XLOG_FMT_LINUX_BE 2
 286 #define XLOG_FMT_IRIX_BE  3
 287
 288 /* our fmt */
 289 #ifdef XFS_NATIVE_HOST
 290 #define XLOG_FMT XLOG_FMT_LINUX_BE
 291 #else
 292 #define XLOG_FMT XLOG_FMT_LINUX_LE
 293 #endif
 294
 295 typedef struct xlog_rec_header {
 296         __be32    h_magicno;    /* log record (LR) identifier           :  4 */
 297         __be32    h_cycle;      /* write cycle of log                   :  4 */
 298         __be32    h_version;    /* LR version                           :  4 */
 299         __be32    h_len;        /* len in bytes; should be 64-bit aligned: 4 */
 300         __be64    h_lsn;        /* lsn of this LR                       :  8 */
 301         __be64    h_tail_lsn;   /* lsn of 1st LR w/ buffers not committed: 8 */
 302         __be32    h_chksum;     /* may not be used; non-zero if used    :  4 */
 303         __be32    h_prev_block; /* block number to previous LR          :  4 */
 304         __be32    h_num_logops; /* number of log operations in this LR  :  4 */
 305         __be32    h_cycle_data[XLOG_HEADER_CYCLE_SIZE / BBSIZE];
 306         /* new fields */
 307         __be32    h_fmt;        /* format of log record                 :  4 */
 308         uuid_t    h_fs_uuid;    /* uuid of FS                           : 16 */
 309         __be32    h_size;       /* iclog size                           :  4 */
 310 } xlog_rec_header_t;
 311
 312 typedef struct xlog_rec_ext_header {
 313         __be32    xh_cycle;     /* write cycle of log                   : 4 */
 314         __be32    xh_cycle_data[XLOG_HEADER_CYCLE_SIZE / BBSIZE]; /*    : 256 */
 315 } xlog_rec_ext_header_t;
 316
 317 #ifdef __KERNEL__
 318
 319 /*
 320  * Quite misnamed, because this union lays out the actual on-disk log buffer.
 321  */
 322 typedef union xlog_in_core2 {
 323         xlog_rec_header_t       hic_header;
 324         xlog_rec_ext_header_t   hic_xheader;
 325         char                    hic_sector[XLOG_HEADER_SIZE];
 326 } xlog_in_core_2_t;
 327
 328 /*
 329  * - A log record header is 512 bytes.  There is plenty of room to grow the
 330  *      xlog_rec_header_t into the reserved space.
 331  * - ic_data follows, so a write to disk can start at the beginning of
 332  *      the iclog.
 333  * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
 334  * - ic_next is the pointer to the next iclog in the ring.
 335  * - ic_bp is a pointer to the buffer used to write this incore log to disk.
 336  * - ic_log is a pointer back to the global log structure.
 337  * - ic_callback is a linked list of callback function/argument pairs to be
 338  *      called after an iclog finishes writing.
 339  * - ic_size is the full size of the header plus data.
 340  * - ic_offset is the current number of bytes written to in this iclog.
 341  * - ic_refcnt is bumped when someone is writing to the log.
 342  * - ic_state is the state of the iclog.
 343  *
 344  * Because of cacheline contention on large machines, we need to separate
 345  * various resources onto different cachelines. To start with, make the
 346  * structure cacheline aligned. The following fields can be contended on
 347  * by independent processes:
 348  *
 349  *      - ic_callback_*
 350  *      - ic_refcnt
 351  *      - fields protected by the global l_icloglock
 352  *
 353  * so we need to ensure that these fields are located in separate cachelines.
 354  * We'll put all the read-only and l_icloglock fields in the first cacheline,
 355  * and move everything else out to subsequent cachelines.
 356  */
 357 typedef struct xlog_in_core {
 358         sv_t                    ic_force_wait;
 359         sv_t                    ic_write_wait;
 360         struct xlog_in_core     *ic_next;
 361         struct xlog_in_core     *ic_prev;
 362         struct xfs_buf          *ic_bp;
 363         struct log              *ic_log;
 364         int                     ic_size;
 365         int                     ic_offset;
 366         int                     ic_bwritecnt;
 367         unsigned short          ic_state;
 368         char                    *ic_datap;      /* pointer to iclog data */
 369
 370         /* Callback structures need their own cacheline */
 371         spinlock_t              ic_callback_lock ____cacheline_aligned_in_smp;
 372         xfs_log_callback_t      *ic_callback;
 373         xfs_log_callback_t      **ic_callback_tail;
 374
 375         /* reference counts need their own cacheline */
 376         atomic_t                ic_refcnt ____cacheline_aligned_in_smp;
 377         xlog_in_core_2_t        *ic_data;
 378 #define ic_header       ic_data->hic_header
 379 } xlog_in_core_t;
 380
 381 /*
 382  * The reservation head lsn is not made up of a cycle number and block number.
 383  * Instead, it uses a cycle number and byte number.  Logs don't expect to
 384  * overflow 31 bits worth of byte offset, so using a byte number will mean
 385  * that round off problems won't occur when releasing partial reservations.
 386  */
 387 typedef struct log {
 388         /* The following fields don't need locking */
 389         struct xfs_mount        *l_mp;          /* mount point */
 390         struct xfs_ail          *l_ailp;        /* AIL log is working with */
 391         struct xfs_buf          *l_xbuf;        /* extra buffer for log
 392                                                  * wrapping */
 393         struct xfs_buftarg      *l_targ;        /* buftarg of log */
 394         uint                    l_flags;
 395         uint                    l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
 396         struct xfs_buf_cancel   **l_buf_cancel_table;
 397         int                     l_iclog_hsize;  /* size of iclog header */
 398         int                     l_iclog_heads;  /* # of iclog header sectors */
 399         uint                    l_sectbb_log;   /* log2 of sector size in BBs */
 400         uint                    l_sectbb_mask;  /* sector size (in BBs)
 401                                                  * alignment mask */
 402         int                     l_iclog_size;   /* size of log in bytes */
 403         int                     l_iclog_size_log; /* log power size of log */
 404         int                     l_iclog_bufs;   /* number of iclog buffers */
 405         xfs_daddr_t             l_logBBstart;   /* start block of log */
 406         int                     l_logsize;      /* size of log in bytes */
 407         int                     l_logBBsize;    /* size of log in BB chunks */
 408
 409         /* The following block of fields are changed while holding icloglock */
 410         sv_t                    l_flush_wait ____cacheline_aligned_in_smp;
 411                                                 /* waiting for iclog flush */
 412         int                     l_covered_state;/* state of "covering disk
 413                                                  * log entries" */
 414         xlog_in_core_t          *l_iclog;       /* head log queue       */
 415         spinlock_t              l_icloglock;    /* grab to change iclog state */
 416         xfs_lsn_t               l_tail_lsn;     /* lsn of 1st LR with unflushed
 417                                                  * buffers */
 418         xfs_lsn_t               l_last_sync_lsn;/* lsn of last LR on disk */
 419         int                     l_curr_cycle;   /* Cycle number of log writes */
 420         int                     l_prev_cycle;   /* Cycle number before last
 421                                                  * block increment */
 422         int                     l_curr_block;   /* current logical log block */
 423         int                     l_prev_block;   /* previous logical log block */
 424
 425         /* The following block of fields are changed while holding grant_lock */
 426         spinlock_t              l_grant_lock ____cacheline_aligned_in_smp;
 427         xlog_ticket_t           *l_reserve_headq;
 428         xlog_ticket_t           *l_write_headq;
 429         int                     l_grant_reserve_cycle;
 430         int                     l_grant_reserve_bytes;
 431         int                     l_grant_write_cycle;
 432         int                     l_grant_write_bytes;
 433
 434         /* The following field are used for debugging; need to hold icloglock */
 435 #ifdef DEBUG
 436         char                    *l_iclog_bak[XLOG_MAX_ICLOGS];
 437 #endif
 438
 439 } xlog_t;
 440
 441 #define XLOG_FORCED_SHUTDOWN(log)       ((log)->l_flags & XLOG_IO_ERROR)
 442
 443
 444 /* common routines */
 445 extern xfs_lsn_t xlog_assign_tail_lsn(struct xfs_mount *mp);
 446 extern int       xlog_recover(xlog_t *log);
 447 extern int       xlog_recover_finish(xlog_t *log);
 448 extern void      xlog_pack_data(xlog_t *log, xlog_in_core_t *iclog, int);
 449
 450 extern kmem_zone_t      *xfs_log_ticket_zone;
 451
 452 /*
 453  * Unmount record type is used as a pseudo transaction type for the ticket.
 454  * It's value must be outside the range of XFS_TRANS_* values.
 455  */
 456 #define XLOG_UNMOUNT_REC_TYPE   (-1U)
 457
 458 #endif  /* __KERNEL__ */
 459
 460 #endif  /* __XFS_LOG_PRIV_H__ */