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