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         long                    expires;
  29         u64                     flush_time;
  30
  31         unsigned                buf_size;       /* size in bytes of @data */
  32         unsigned                sectors;        /* maximum size for current entry */
  33         unsigned                disk_sectors;   /* maximum size entry could have been, if
  34                                                    buf_size was bigger */
  35         unsigned                u64s_reserved;
  36         bool                    noflush;        /* write has already been kicked off, and was noflush */
  37         bool                    must_flush;     /* something wants a flush */
  38         bool                    separate_flush;
  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 enum journal_pin_type {
  47         JOURNAL_PIN_btree,
  48         JOURNAL_PIN_key_cache,
  49         JOURNAL_PIN_other,
  50         JOURNAL_PIN_NR,
  51 };
  52
  53 struct journal_entry_pin_list {
  54         struct list_head                list[JOURNAL_PIN_NR];
  55         struct list_head                flushed;
  56         atomic_t                        count;
  57         struct bch_devs_list            devs;
  58 };
  59
  60 struct journal;
  61 struct journal_entry_pin;
  62 typedef int (*journal_pin_flush_fn)(struct journal *j,
  63                                 struct journal_entry_pin *, u64);
  64
  65 struct journal_entry_pin {
  66         struct list_head                list;
  67         journal_pin_flush_fn            flush;
  68         u64                             seq;
  69 };
  70
  71 struct journal_res {
  72         bool                    ref;
  73         u8                      idx;
  74         u16                     u64s;
  75         u32                     offset;
  76         u64                     seq;
  77 };
  78
  79 union journal_res_state {
  80         struct {
  81                 atomic64_t      counter;
  82         };
  83
  84         struct {
  85                 u64             v;
  86         };
  87
  88         struct {
  89                 u64             cur_entry_offset:20,
  90                                 idx:2,
  91                                 unwritten_idx:2,
  92                                 buf0_count:10,
  93                                 buf1_count:10,
  94                                 buf2_count:10,
  95                                 buf3_count:10;
  96         };
  97 };
  98
  99 /* bytes: */
 100 #define JOURNAL_ENTRY_SIZE_MIN          (64U << 10) /* 64k */
 101 #define JOURNAL_ENTRY_SIZE_MAX          (4U  << 20) /* 4M */
 102
 103 /*
 104  * We stash some journal state as sentinal values in cur_entry_offset:
 105  * note - cur_entry_offset is in units of u64s
 106  */
 107 #define JOURNAL_ENTRY_OFFSET_MAX        ((1U << 20) - 1)
 108
 109 #define JOURNAL_ENTRY_CLOSED_VAL        (JOURNAL_ENTRY_OFFSET_MAX - 1)
 110 #define JOURNAL_ENTRY_ERROR_VAL         (JOURNAL_ENTRY_OFFSET_MAX)
 111
 112 struct journal_space {
 113         /* Units of 512 bytes sectors: */
 114         unsigned        next_entry; /* How big the next journal entry can be */
 115         unsigned        total;
 116 };
 117
 118 enum journal_space_from {
 119         journal_space_discarded,
 120         journal_space_clean_ondisk,
 121         journal_space_clean,
 122         journal_space_total,
 123         journal_space_nr,
 124 };
 125
 126 enum journal_flags {
 127         JOURNAL_REPLAY_DONE,
 128         JOURNAL_STARTED,
 129         JOURNAL_MAY_SKIP_FLUSH,
 130         JOURNAL_NEED_FLUSH_WRITE,
 131 };
 132
 133 /* Reasons we may fail to get a journal reservation: */
 134 #define JOURNAL_ERRORS()                \
 135         x(ok)                           \
 136         x(blocked)                      \
 137         x(max_in_flight)                \
 138         x(journal_full)                 \
 139         x(journal_pin_full)             \
 140         x(journal_stuck)                \
 141         x(insufficient_devices)
 142
 143 enum journal_errors {
 144 #define x(n)    JOURNAL_ERR_##n,
 145         JOURNAL_ERRORS()
 146 #undef x
 147 };
 148
 149 typedef DARRAY(u64)             darray_u64;
 150
 151 /* Embedded in struct bch_fs */
 152 struct journal {
 153         /* Fastpath stuff up front: */
 154         struct {
 155
 156         union journal_res_state reservations;
 157         enum bch_watermark      watermark;
 158
 159         } __aligned(SMP_CACHE_BYTES);
 160
 161         unsigned long           flags;
 162
 163         /* Max size of current journal entry */
 164         unsigned                cur_entry_u64s;
 165         unsigned                cur_entry_sectors;
 166
 167         /* Reserved space in journal entry to be used just prior to write */
 168         unsigned                entry_u64s_reserved;
 169
 170
 171         /*
 172          * 0, or -ENOSPC if waiting on journal reclaim, or -EROFS if
 173          * insufficient devices:
 174          */
 175         enum journal_errors     cur_entry_error;
 176
 177         unsigned                buf_size_want;
 178         /*
 179          * We may queue up some things to be journalled (log messages) before
 180          * the journal has actually started - stash them here:
 181          */
 182         darray_u64              early_journal_entries;
 183
 184         /*
 185          * Two journal entries -- one is currently open for new entries, the
 186          * other is possibly being written out.
 187          */
 188         struct journal_buf      buf[JOURNAL_BUF_NR];
 189
 190         spinlock_t              lock;
 191
 192         /* if nonzero, we may not open a new journal entry: */
 193         unsigned                blocked;
 194
 195         /* Used when waiting because the journal was full */
 196         wait_queue_head_t       wait;
 197         struct closure_waitlist async_wait;
 198         struct closure_waitlist preres_wait;
 199
 200         struct closure          io;
 201         struct delayed_work     write_work;
 202
 203         /* Sequence number of most recent journal entry (last entry in @pin) */
 204         atomic64_t              seq;
 205
 206         /* seq, last_seq from the most recent journal entry successfully written */
 207         u64                     seq_ondisk;
 208         u64                     flushed_seq_ondisk;
 209         u64                     last_seq_ondisk;
 210         u64                     err_seq;
 211         u64                     last_empty_seq;
 212
 213         /*
 214          * FIFO of journal entries whose btree updates have not yet been
 215          * written out.
 216          *
 217          * Each entry is a reference count. The position in the FIFO is the
 218          * entry's sequence number relative to @seq.
 219          *
 220          * The journal entry itself holds a reference count, put when the
 221          * journal entry is written out. Each btree node modified by the journal
 222          * entry also holds a reference count, put when the btree node is
 223          * written.
 224          *
 225          * When a reference count reaches zero, the journal entry is no longer
 226          * needed. When all journal entries in the oldest journal bucket are no
 227          * longer needed, the bucket can be discarded and reused.
 228          */
 229         struct {
 230                 u64 front, back, size, mask;
 231                 struct journal_entry_pin_list *data;
 232         }                       pin;
 233
 234         struct journal_space    space[journal_space_nr];
 235
 236         u64                     replay_journal_seq;
 237         u64                     replay_journal_seq_end;
 238
 239         struct write_point      wp;
 240         spinlock_t              err_lock;
 241
 242         struct mutex            reclaim_lock;
 243         /*
 244          * Used for waiting until journal reclaim has freed up space in the
 245          * journal:
 246          */
 247         wait_queue_head_t       reclaim_wait;
 248         struct task_struct      *reclaim_thread;
 249         bool                    reclaim_kicked;
 250         unsigned long           next_reclaim;
 251         u64                     nr_direct_reclaim;
 252         u64                     nr_background_reclaim;
 253
 254         unsigned long           last_flushed;
 255         struct journal_entry_pin *flush_in_progress;
 256         bool                    flush_in_progress_dropped;
 257         wait_queue_head_t       pin_flush_wait;
 258
 259         /* protects advancing ja->discard_idx: */
 260         struct mutex            discard_lock;
 261         bool                    can_discard;
 262
 263         unsigned long           last_flush_write;
 264
 265         u64                     write_start_time;
 266
 267         u64                     nr_flush_writes;
 268         u64                     nr_noflush_writes;
 269         u64                     entry_bytes_written;
 270
 271         u64                     low_on_space_start;
 272         u64                     low_on_pin_start;
 273         u64                     max_in_flight_start;
 274
 275         struct bch2_time_stats  *flush_write_time;
 276         struct bch2_time_stats  *noflush_write_time;
 277         struct bch2_time_stats  *flush_seq_time;
 278
 279 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 280         struct lockdep_map      res_map;
 281 #endif
 282 } __aligned(SMP_CACHE_BYTES);
 283
 284 /*
 285  * Embedded in struct bch_dev. First three fields refer to the array of journal
 286  * buckets, in bch_sb.
 287  */
 288 struct journal_device {
 289         /*
 290          * For each journal bucket, contains the max sequence number of the
 291          * journal writes it contains - so we know when a bucket can be reused.
 292          */
 293         u64                     *bucket_seq;
 294
 295         unsigned                sectors_free;
 296
 297         /*
 298          * discard_idx <= dirty_idx_ondisk <= dirty_idx <= cur_idx:
 299          */
 300         unsigned                discard_idx;            /* Next bucket to discard */
 301         unsigned                dirty_idx_ondisk;
 302         unsigned                dirty_idx;
 303         unsigned                cur_idx;                /* Journal bucket we're currently writing to */
 304         unsigned                nr;
 305
 306         u64                     *buckets;
 307
 308         /* Bio for journal reads/writes to this device */
 309         struct bio              *bio;
 310
 311         /* for bch_journal_read_device */
 312         struct closure          read;
 313 };
 314
 315 /*
 316  * journal_entry_res - reserve space in every journal entry:
 317  */
 318 struct journal_entry_res {
 319         unsigned                u64s;
 320 };
 321
 322 #endif /* _BCACHEFS_JOURNAL_TYPES_H */