git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/journal_types.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _BCACHEFS_JOURNAL_TYPES_H
   3 #define _BCACHEFS_JOURNAL_TYPES_H
   4
   5 #include <linux/cache.h>
   6 #include <linux/workqueue.h>
   7
   8 #include "alloc_types.h"
   9 #include "super_types.h"
  10 #include "fifo.h"
  11
  12 #define JOURNAL_BUF_BITS        2
  13 #define JOURNAL_BUF_NR          (1U << JOURNAL_BUF_BITS)
  14 #define JOURNAL_BUF_MASK        (JOURNAL_BUF_NR - 1)
  15
  16 /*
  17  * We put JOURNAL_BUF_NR of these in struct journal; we used them for writes to
  18  * the journal that are being staged or in flight.
  19  */
  20 struct journal_buf {
  21         struct jset             *data;
  22
  23         __BKEY_PADDED(key, BCH_REPLICAS_MAX);
  24
  25         struct closure_waitlist wait;
  26         u64                     last_seq;       /* copy of data->last_seq */
  27
  28         unsigned                buf_size;       /* size in bytes of @data */
  29         unsigned                sectors;        /* maximum size for current entry */
  30         unsigned                disk_sectors;   /* maximum size entry could have been, if
  31                                                    buf_size was bigger */
  32         unsigned                u64s_reserved;
  33         bool                    noflush;        /* write has already been kicked off, and was noflush */
  34         bool                    must_flush;     /* something wants a flush */
  35         bool                    separate_flush;
  36         /* bloom filter: */
  37         unsigned long           has_inode[1024 / sizeof(unsigned long)];
  38 };
  39
  40 /*
  41  * Something that makes a journal entry dirty - i.e. a btree node that has to be
  42  * flushed:
  43  */
  44
  45 struct journal_entry_pin_list {
  46         struct list_head                list;
  47         struct list_head                key_cache_list;
  48         struct list_head                flushed;
  49         atomic_t                        count;
  50         struct bch_devs_list            devs;
  51 };
  52
  53 struct journal;
  54 struct journal_entry_pin;
  55 typedef int (*journal_pin_flush_fn)(struct journal *j,
  56                                 struct journal_entry_pin *, u64);
  57
  58 struct journal_entry_pin {
  59         struct list_head                list;
  60         journal_pin_flush_fn            flush;
  61         u64                             seq;
  62 };
  63
  64 struct journal_res {
  65         bool                    ref;
  66         u8                      idx;
  67         u16                     u64s;
  68         u32                     offset;
  69         u64                     seq;
  70 };
  71
  72 /*
  73  * For reserving space in the journal prior to getting a reservation on a
  74  * particular journal entry:
  75  */
  76 struct journal_preres {
  77         unsigned                u64s;
  78 };
  79
  80 union journal_res_state {
  81         struct {
  82                 atomic64_t      counter;
  83         };
  84
  85         struct {
  86                 u64             v;
  87         };
  88
  89         struct {
  90                 u64             cur_entry_offset:20,
  91                                 idx:2,
  92                                 unwritten_idx:2,
  93                                 buf0_count:10,
  94                                 buf1_count:10,
  95                                 buf2_count:10,
  96                                 buf3_count:10;
  97         };
  98 };
  99
 100 union journal_preres_state {
 101         struct {
 102                 atomic64_t      counter;
 103         };
 104
 105         struct {
 106                 u64             v;
 107         };
 108
 109         struct {
 110                 u64             waiting:1,
 111                                 reserved:31,
 112                                 remaining:32;
 113         };
 114 };
 115
 116 /* bytes: */
 117 #define JOURNAL_ENTRY_SIZE_MIN          (64U << 10) /* 64k */
 118 #define JOURNAL_ENTRY_SIZE_MAX          (4U  << 20) /* 4M */
 119
 120 /*
 121  * We stash some journal state as sentinal values in cur_entry_offset:
 122  * note - cur_entry_offset is in units of u64s
 123  */
 124 #define JOURNAL_ENTRY_OFFSET_MAX        ((1U << 20) - 1)
 125
 126 #define JOURNAL_ENTRY_CLOSED_VAL        (JOURNAL_ENTRY_OFFSET_MAX - 1)
 127 #define JOURNAL_ENTRY_ERROR_VAL         (JOURNAL_ENTRY_OFFSET_MAX)
 128
 129 struct journal_space {
 130         /* Units of 512 bytes sectors: */
 131         unsigned        next_entry; /* How big the next journal entry can be */
 132         unsigned        total;
 133 };
 134
 135 enum journal_space_from {
 136         journal_space_discarded,
 137         journal_space_clean_ondisk,
 138         journal_space_clean,
 139         journal_space_total,
 140         journal_space_nr,
 141 };
 142
 143 /*
 144  * JOURNAL_NEED_WRITE - current (pending) journal entry should be written ASAP,
 145  * either because something's waiting on the write to complete or because it's
 146  * been dirty too long and the timer's expired.
 147  */
 148
 149 enum {
 150         JOURNAL_REPLAY_DONE,
 151         JOURNAL_STARTED,
 152         JOURNAL_RECLAIM_STARTED,
 153         JOURNAL_NEED_WRITE,
 154         JOURNAL_MAY_GET_UNRESERVED,
 155         JOURNAL_MAY_SKIP_FLUSH,
 156 };
 157
 158 /* Embedded in struct bch_fs */
 159 struct journal {
 160         /* Fastpath stuff up front: */
 161
 162         unsigned long           flags;
 163
 164         union journal_res_state reservations;
 165
 166         /* Max size of current journal entry */
 167         unsigned                cur_entry_u64s;
 168         unsigned                cur_entry_sectors;
 169
 170         /*
 171          * 0, or -ENOSPC if waiting on journal reclaim, or -EROFS if
 172          * insufficient devices:
 173          */
 174         enum {
 175                 cur_entry_ok,
 176                 cur_entry_blocked,
 177                 cur_entry_journal_full,
 178                 cur_entry_journal_pin_full,
 179                 cur_entry_journal_stuck,
 180                 cur_entry_insufficient_devices,
 181         }                       cur_entry_error;
 182
 183         union journal_preres_state prereserved;
 184
 185         /* Reserved space in journal entry to be used just prior to write */
 186         unsigned                entry_u64s_reserved;
 187
 188         unsigned                buf_size_want;
 189
 190         /*
 191          * Two journal entries -- one is currently open for new entries, the
 192          * other is possibly being written out.
 193          */
 194         struct journal_buf      buf[JOURNAL_BUF_NR];
 195
 196         spinlock_t              lock;
 197
 198         /* if nonzero, we may not open a new journal entry: */
 199         unsigned                blocked;
 200
 201         /* Used when waiting because the journal was full */
 202         wait_queue_head_t       wait;
 203         struct closure_waitlist async_wait;
 204         struct closure_waitlist preres_wait;
 205
 206         struct closure          io;
 207         struct delayed_work     write_work;
 208
 209         /* Sequence number of most recent journal entry (last entry in @pin) */
 210         atomic64_t              seq;
 211
 212         /* seq, last_seq from the most recent journal entry successfully written */
 213         u64                     seq_ondisk;
 214         u64                     flushed_seq_ondisk;
 215         u64                     last_seq_ondisk;
 216         u64                     err_seq;
 217         u64                     last_empty_seq;
 218
 219         /*
 220          * FIFO of journal entries whose btree updates have not yet been
 221          * written out.
 222          *
 223          * Each entry is a reference count. The position in the FIFO is the
 224          * entry's sequence number relative to @seq.
 225          *
 226          * The journal entry itself holds a reference count, put when the
 227          * journal entry is written out. Each btree node modified by the journal
 228          * entry also holds a reference count, put when the btree node is
 229          * written.
 230          *
 231          * When a reference count reaches zero, the journal entry is no longer
 232          * needed. When all journal entries in the oldest journal bucket are no
 233          * longer needed, the bucket can be discarded and reused.
 234          */
 235         struct {
 236                 u64 front, back, size, mask;
 237                 struct journal_entry_pin_list *data;
 238         }                       pin;
 239
 240         struct journal_space    space[journal_space_nr];
 241
 242         u64                     replay_journal_seq;
 243         u64                     replay_journal_seq_end;
 244
 245         struct write_point      wp;
 246         spinlock_t              err_lock;
 247
 248         struct mutex            reclaim_lock;
 249         wait_queue_head_t       reclaim_wait;
 250         struct task_struct      *reclaim_thread;
 251         bool                    reclaim_kicked;
 252         unsigned long           next_reclaim;
 253         u64                     nr_direct_reclaim;
 254         u64                     nr_background_reclaim;
 255
 256         unsigned long           last_flushed;
 257         struct journal_entry_pin *flush_in_progress;
 258         bool                    flush_in_progress_dropped;
 259         wait_queue_head_t       pin_flush_wait;
 260
 261         /* protects advancing ja->discard_idx: */
 262         struct mutex            discard_lock;
 263         bool                    can_discard;
 264
 265         unsigned                write_delay_ms;
 266         unsigned                reclaim_delay_ms;
 267         unsigned long           last_flush_write;
 268
 269         u64                     res_get_blocked_start;
 270         u64                     need_write_time;
 271         u64                     write_start_time;
 272
 273         u64                     nr_flush_writes;
 274         u64                     nr_noflush_writes;
 275
 276         struct time_stats       *write_time;
 277         struct time_stats       *delay_time;
 278         struct time_stats       *blocked_time;
 279         struct time_stats       *flush_seq_time;
 280
 281 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 282         struct lockdep_map      res_map;
 283 #endif
 284 };
 285
 286 /*
 287  * Embedded in struct bch_dev. First three fields refer to the array of journal
 288  * buckets, in bch_sb.
 289  */
 290 struct journal_device {
 291         /*
 292          * For each journal bucket, contains the max sequence number of the
 293          * journal writes it contains - so we know when a bucket can be reused.
 294          */
 295         u64                     *bucket_seq;
 296
 297         unsigned                sectors_free;
 298
 299         /*
 300          * discard_idx <= dirty_idx_ondisk <= dirty_idx <= cur_idx:
 301          */
 302         unsigned                discard_idx;            /* Next bucket to discard */
 303         unsigned                dirty_idx_ondisk;
 304         unsigned                dirty_idx;
 305         unsigned                cur_idx;                /* Journal bucket we're currently writing to */
 306         unsigned                nr;
 307
 308         u64                     *buckets;
 309
 310         /* Bio for journal reads/writes to this device */
 311         struct bio              *bio;
 312
 313         /* for bch_journal_read_device */
 314         struct closure          read;
 315 };
 316
 317 /*
 318  * journal_entry_res - reserve space in every journal entry:
 319  */
 320 struct journal_entry_res {
 321         unsigned                u64s;
 322 };
 323
 324 #endif /* _BCACHEFS_JOURNAL_TYPES_H */