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