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

   1 #ifndef _BCACHEFS_JOURNAL_TYPES_H
   2 #define _BCACHEFS_JOURNAL_TYPES_H
   3
   4 #include <linux/cache.h>
   5 #include <linux/workqueue.h>
   6
   7 #include "alloc_types.h"
   8 #include "super_types.h"
   9 #include "fifo.h"
  10
  11 struct journal_res;
  12
  13 /*
  14  * We put two of these in struct journal; we used them for writes to the
  15  * journal that are being staged or in flight.
  16  */
  17 struct journal_buf {
  18         struct jset             *data;
  19
  20         BKEY_PADDED(key);
  21
  22         struct closure_waitlist wait;
  23
  24         unsigned                size;
  25         unsigned                disk_sectors;
  26         /* bloom filter: */
  27         unsigned long           has_inode[1024 / sizeof(unsigned long)];
  28 };
  29
  30 /*
  31  * Something that makes a journal entry dirty - i.e. a btree node that has to be
  32  * flushed:
  33  */
  34
  35 struct journal_entry_pin_list {
  36         struct list_head                list;
  37         struct list_head                flushed;
  38         atomic_t                        count;
  39         struct bch_devs_list            devs;
  40 };
  41
  42 struct journal;
  43 struct journal_entry_pin;
  44 typedef void (*journal_pin_flush_fn)(struct journal *j,
  45                                 struct journal_entry_pin *, u64);
  46
  47 struct journal_entry_pin {
  48         struct list_head                list;
  49         journal_pin_flush_fn            flush;
  50         struct journal_entry_pin_list   *pin_list;
  51 };
  52
  53 /* corresponds to a btree node with a blacklisted bset: */
  54 struct blacklisted_node {
  55         __le64                  seq;
  56         enum btree_id           btree_id;
  57         struct bpos             pos;
  58 };
  59
  60 struct journal_seq_blacklist {
  61         struct list_head        list;
  62         u64                     seq;
  63         bool                    written;
  64         struct journal_entry_pin pin;
  65
  66         struct blacklisted_node *entries;
  67         size_t                  nr_entries;
  68 };
  69
  70 struct journal_res {
  71         bool                    ref;
  72         u8                      idx;
  73         u16                     u64s;
  74         u32                     offset;
  75         u64                     seq;
  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:1,
  90                                 prev_buf_unwritten:1,
  91                                 buf0_count:21,
  92                                 buf1_count:21;
  93         };
  94 };
  95
  96 /* bytes: */
  97 #define JOURNAL_ENTRY_SIZE_MIN          (64U << 10) /* 64k */
  98 #define JOURNAL_ENTRY_SIZE_MAX          (4U  << 20) /* 4M */
  99
 100 /*
 101  * We stash some journal state as sentinal values in cur_entry_offset:
 102  * note - cur_entry_offset is in units of u64s
 103  */
 104 #define JOURNAL_ENTRY_OFFSET_MAX        ((1U << 20) - 1)
 105
 106 #define JOURNAL_ENTRY_CLOSED_VAL        (JOURNAL_ENTRY_OFFSET_MAX - 1)
 107 #define JOURNAL_ENTRY_ERROR_VAL         (JOURNAL_ENTRY_OFFSET_MAX)
 108
 109 /*
 110  * JOURNAL_NEED_WRITE - current (pending) journal entry should be written ASAP,
 111  * either because something's waiting on the write to complete or because it's
 112  * been dirty too long and the timer's expired.
 113  */
 114
 115 enum {
 116         JOURNAL_REPLAY_DONE,
 117         JOURNAL_STARTED,
 118         JOURNAL_NEED_WRITE,
 119 };
 120
 121 /* Embedded in struct bch_fs */
 122 struct journal {
 123         /* Fastpath stuff up front: */
 124
 125         unsigned long           flags;
 126
 127         union journal_res_state reservations;
 128         unsigned                cur_entry_u64s;
 129         unsigned                prev_buf_sectors;
 130         unsigned                cur_buf_sectors;
 131         unsigned                buf_size_want;
 132
 133         /*
 134          * Two journal entries -- one is currently open for new entries, the
 135          * other is possibly being written out.
 136          */
 137         struct journal_buf      buf[2];
 138
 139         spinlock_t              lock;
 140
 141         /* Used when waiting because the journal was full */
 142         wait_queue_head_t       wait;
 143         struct closure_waitlist async_wait;
 144
 145         struct closure          io;
 146         struct delayed_work     write_work;
 147
 148         /* Sequence number of most recent journal entry (last entry in @pin) */
 149         atomic64_t              seq;
 150
 151         /* last_seq from the most recent journal entry written */
 152         u64                     last_seq_ondisk;
 153
 154         /*
 155          * FIFO of journal entries whose btree updates have not yet been
 156          * written out.
 157          *
 158          * Each entry is a reference count. The position in the FIFO is the
 159          * entry's sequence number relative to @seq.
 160          *
 161          * The journal entry itself holds a reference count, put when the
 162          * journal entry is written out. Each btree node modified by the journal
 163          * entry also holds a reference count, put when the btree node is
 164          * written.
 165          *
 166          * When a reference count reaches zero, the journal entry is no longer
 167          * needed. When all journal entries in the oldest journal bucket are no
 168          * longer needed, the bucket can be discarded and reused.
 169          */
 170         struct {
 171                 u64 front, back, size, mask;
 172                 struct journal_entry_pin_list *data;
 173         }                       pin;
 174         struct journal_entry_pin_list *replay_pin_list;
 175
 176         struct mutex            blacklist_lock;
 177         struct list_head        seq_blacklist;
 178
 179         BKEY_PADDED(key);
 180         struct write_point      wp;
 181         spinlock_t              err_lock;
 182
 183         struct delayed_work     reclaim_work;
 184         unsigned long           last_flushed;
 185
 186         /* protects advancing ja->last_idx: */
 187         struct mutex            reclaim_lock;
 188         unsigned                write_delay_ms;
 189         unsigned                reclaim_delay_ms;
 190
 191         u64                     res_get_blocked_start;
 192         u64                     need_write_time;
 193         u64                     write_start_time;
 194
 195         struct time_stats       *write_time;
 196         struct time_stats       *delay_time;
 197         struct time_stats       *blocked_time;
 198         struct time_stats       *flush_seq_time;
 199
 200 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 201         struct lockdep_map      res_map;
 202 #endif
 203 };
 204
 205 /*
 206  * Embedded in struct bch_dev. First three fields refer to the array of journal
 207  * buckets, in bch_sb.
 208  */
 209 struct journal_device {
 210         /*
 211          * For each journal bucket, contains the max sequence number of the
 212          * journal writes it contains - so we know when a bucket can be reused.
 213          */
 214         u64                     *bucket_seq;
 215
 216         unsigned                sectors_free;
 217
 218         /* Journal bucket we're currently writing to */
 219         unsigned                cur_idx;
 220
 221         /* Last journal bucket that still contains an open journal entry */
 222
 223         /*
 224          * j->lock and j->reclaim_lock must both be held to modify, j->lock
 225          * sufficient to read:
 226          */
 227         unsigned                last_idx;
 228         unsigned                nr;
 229         u64                     *buckets;
 230
 231         /* Bio for journal reads/writes to this device */
 232         struct bio              *bio;
 233
 234         /* for bch_journal_read_device */
 235         struct closure          read;
 236 };
 237
 238 #endif /* _BCACHEFS_JOURNAL_TYPES_H */